







































NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 
93-126 

93 

 

 

 

Review 

GI meets AI: Glycemic index in the age of AI, 

computational breakthroughs 
N.H. Wanigasingha1*, W.G.L. Harshani1, M.K.A. Ariyaratne1, T.G.I. Fernando1, U. Dikwatta1, U.S. 

Samarasinghe2 

1Department of Computer Science, Faculty of Applied Sciences, University of Sri Jayewardenepura, Sri Lanka 
2Department of Information Technology, Faculty of Management Studies and Commerece, University of Sri 

Jayewardenepura, Sri Lanka 

A R T I C L E   I N F O 
 

Article history: 
Received 24 August 2025  
Received in revised form 
10 October 2025 
Accepted 26 October 2025 
 
Keywords:  
Artificial intelligence, Glycemic index, 
Machine learning, Deep learning, 
Food analytics 
 
*Corresponding author 
Email address: 
hirushanethni@sjp.ac.lk 
 
DOI: 10.55670/fpll.futech.5.1.10 

A B S T R A C T 
 

This review explores the avenues for the application of Artificial Intelligence 
(AI) techniques in Glycemic Index (GI) related research. The necessity of 
sophisticated technologies to investigate various GI‐related studies in food 
analytics has been established in recent years. AI technologies have emerged as 
promising approaches to address these challenges. We identified six major AI 
technologies applied in GI research: Machine Learning, Reinforcement 
Learning, Deep Learning, Image Processing, Natural Language Processing, and 
Explainable AI. Some of our findings include: (a) There have been significant 
improvements in GI-related studies using AI technologies over the past decade. 
(b) Machine learning algorithms were widely used (c) Many researchers used 
custom datasets, with the predominance of research originating from North 
American countries. (d) Identification of limitations and future directions for 
GI‐related studies employing AI technologies. By embracing AI technologies, the 
field of food analytics is poised for substantial advancements in understanding 
and managing glycemic responses. Unlike existing reviews that mainly discuss 
nutritional or clinical aspects of the glycemic index, this study systematically 
examines the integration of AI and machine learning technologies in GI-related 
research. It highlights computational breakthroughs, methodological trends, 
and future directions for intelligent glycemic analysis. 

1. Introduction 

Of all the challenges technology seeks to address, human 
health stands as the most critical and universally compelling. 
Recent evidence shows that Artificial Intelligence is having a 
significant impact on the healthcare industry, highlighting 
how important human health has become for technological 
advancements. In 2024, a survey by the Berkeley Research 
Group found that healthcare providers and pharmaceutical 
professionals are increasingly relying on AI to enhance 
patient care, streamline processes, and transform the delivery 
of medical treatment [1]. These breakthroughs are 
particularly evident in diagnostics and personalized 
treatments, where AI has shown impressive precision. For 
example, AI models can now assess cancer aggressiveness 
more accurately than traditional biopsies, according to a 2024 
report from the World Economic Forum [2]. At the same time, 
the global market for AI‐driven healthcare is expected to 
reach $70 billion by 2032, fueled by advances in AI for drug 
discovery and medical imaging. In countries like China, AI 
plays a vital role in optimizing medical resource distribution 
and enhancing diagnostic accuracy, especially in areas with 
limited access to healthcare [3]. Additionally, reviews of AI 

healthcare studies from 2023 show that fields like radiology 
and Gastroenterology are experiencing the greatest impact. 
Looking ahead, experts predict that AI will have a wider 
influence across many areas of medicine, including 
administration and education [4]. Among the leading 
noncommunicable diseases, diabetes is now one of the 
primary causes of death worldwide, representing an 
escalating global health threat [5]. Therefore, extensive 
research has been conducted, and active research is currently 
underway, to find solutions to prevent and manage it [6-8]. 
Diabetes is a chronic, noncommunicable disease that occurs 
when the body is either unable to produce enough insulin or 
cannot effectively use the insulin it produces. Insulin is a 
hormone that regulates blood sugar (glucose) levels, which is 
crucial for providing energy to the body’s cells. There are two 
main types of diabetes:  
• Type 1 diabetes: This form is often diagnosed in children 

and young adults, though it can occur at any age. It happens 
when the immune system mistakenly attacks and destroys 
the insulin-producing beta cells in the pancreas, leading to 
little or no insulin production. People with Type 1 diabetes 
require lifelong insulin therapy [9]. 

 

 

Future Technology 

Open Access Journal 

https://doi.org/10.55670/fpll.futech.5.1.10 

 

 

February 2026| Volume 05 | Issue 01 | Pages 93-126 

Journal homepage: https://fupubco.com/futech 

 

ISSN 2832-0379 

mailto:hirushanethni@sjp.ac.lk
https://doi.org/10.55670/fpll.futech.5.1.10
https://fupubco.com/futech


NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

94 

 

 
• Type 2 diabetes: This is the most common form and is often 

linked to lifestyle factors such as obesity, poor diet, and 
physical inactivity. In this condition, the body becomes 
resistant to insulin, or the pancreas cannot produce enough 
insulin to maintain normal blood glucose levels. Type 2 
diabetes can often be managed with lifestyle changes, but 
may also require medication or insulin [9]. 

Diabetes develops when the glucose in the blood remains 
elevated over time, leading to serious health complications, 
such as heart disease, kidney damage, nerve damage, and 
vision problems. Early detection and management through 
lifestyle changes, medication, and regular monitoring of blood 
sugar levels are crucial in preventing or delaying these 
complications. There are several factors, such as Glycemic 
Index, Glycemic Load, Fiber Content, Carbohydrate Type, 
Meal Timing and Composition, Physical Activity, and others, 
that influence blood glucose levels and diabetes. Among them, 
the Glycemic Index plays a key role. The Glycemic Index is a 
scale that measures how fast the carbohydrates in different 
foods raise your blood sugar after you eat them. Foods are 
ranked from 0 to 100, with higher numbers meaning they 
cause a quicker rise in blood glucose, while lower numbers 
indicate a slower, steadier increase. 
• Low GI foods (GI ≤ 55): These cause a gradual rise in blood 

sugar, helping maintain stable levels (e.g., most fruits, 
vegetables, whole grains, and legumes). 

• Medium GI foods (GI 56–69): These create a moderate 
increase in blood glucose (e.g., rye bread, bananas, sweet 
potatoes). 

• High GI foods (GI ≥ 70): These lead to rapid spikes in blood 
sugar (e.g., white bread, sugary drinks, and processed 
cereals). 

For people with diabetes, keeping blood sugar levels 
under control is essential to managing the condition and 
preventing complications. Eating high‐GI foods can cause 
sudden blood sugar spikes, which can be dangerous for 
diabetics who may struggle with insulin production or use. On 
the other hand, low‐GI foods help keep blood sugar stable, 
making it easier to manage diabetes. The GI is especially 
important in Type 2 diabetes, where lifestyle and dietary 
choices play a huge role. By choosing lower‐GI foods, people 
with diabetes can prevent sharp rises in blood sugar, reducing 
the need for insulin and supporting better long‐term blood 
sugar control. Due to the importance of the GI in the 
management of blood sugar levels, especially for people with 
diabetes, extensive research has been conducted to explore 
its various applications. Research related to GI has been 

observed to grow rapidly over the past few decades (Figure 
1). Studies have connected GI with various fields, 
investigating its role in diet, health outcomes, and disease 
management. Researchers have examined how different 
types of foods affect blood glucose levels, developed 
predictive models for GI, and explored the benefits of a low‐
GI diet in preventing and managing diabetes, obesity, and 
cardiovascular disease. These efforts aim to deepen the 
understanding of the significance of GI and provide actionable 
insights to improve health and wellness. 

 
Figure 1. Number of papers by year 

It can be quite challenging to determine which 
technologies are best suited for different applications of the 
GI and to understand the reasons behind their effectiveness. 
Whereas a few prior reviews investigated the GI and nutrition 
science/diabetes research, many of these centered on 
clinical/dietary applications and did not systematically 
examine AI/ML contributions in this field. When multiple 
technologies address the same problem, it becomes even 
more important to make comparisons. This highlights the 
need for a systematic review and analysis. The primary goal 
of this paper is to gather and examine various studies where 
different approaches have been applied to the GI, with the aim 
of uncovering useful insights related to human health. By 
reviewing these approaches in a structured manner, we aim 
to achieve two main objectives: first, to present and analyze 
the areas where different technologies have been successfully 
used with the GI, particularly in predicting GI values; and 
second, to expand the potential applications of GI for a wider 
audience. 

Here, we provide a comprehensive overview of 
technologies like deep learning, reinforcement learning, 
explainable AI, and natural language processing, and how 
these technologies are utilized to combine viewpoints of food 
analytics and computational intelligence to predict GI. Unlike 
prior reviews, which merely described dietary effects, our 
work systematically charted how AI methodologies evolved 
for predicting, monitoring, and personalizing nutrition plans 
for GI. Moreover, our review identifies areas of poor usage of 
state-of-the-art AI paradigms (e.g., XAI, multimodal data 
integration) and offers future research prospects for 
minimizing the gap between food science and computational 
intelligence. 

To enhance the reader’s experience, the rest of this paper 
is structured as follows: in Section 2, we outline the 
methodology used to conduct the review. In the next section, 
we will provide a brief overview of these technologies, 
offering the reader a foundational understanding of concepts 
such as Computer Vision, Deep Learning, Food Technologies, 
Image Processing, Machine Learning, Reinforcement 

Abbreviations 

AI  Artificial Intelligence  

ANN  Artificial Neural Network 

CGM  Continuous Glucose Monitoring 

DL  Deep Learning 

GL  Glycemic Load 

GI   Glycemic Index  

HbA1c  Glycated Hemoglobin 

IoT  Internet of Things 

ML  Machine Learning 

NLP  Natural Language Processing 

PRISMA Preferred Reporting Items for Systematic 

Reviews and Meta‐Analyses  

RL   Reinforcement Learning 

SHAP  SHapley Additive exPlanations  

XAI  Explainable Artificial Intelligence 

IP                      ImageProcessing 

 

 

 

 

 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

95 

 

Learning, Natural Language Processing, Statistical Analysis, 
and Mathematical Modeling. The research studies, divided 
into those key concepts, are discussed in Section 4, all related 
to the use of GI and Information Technology. In Section 5, we 
discuss the standard datasets used in the selected studies. 
Finally, we conclude the review with a short discussion. 

2. Methodology for systematic review: applying the 

PRISMA framework 

To enhance the transparency of our research reporting, 
this systematic review was conducted in accordance with the 
Preferred Reporting Items for systematic reviews and Meta‐
Analyses (PRISMA) guidelines, ensuring a comprehensive 
approach to the review [10]. In our initial stages of research, 
we began by identifying relevant literature using the key 
terms ‘Glycemic Index’ and ‘machine learning’ to gather 
foundational insights. In addition to these, we expanded our 
search scope by incorporating various AI and data prediction‐
driven keywords such as ‘deep learning’, ‘NLP’, ‘data science’, 
‘machine learning’, ‘reinforcement learning’, and ‘statistical 
mathematics’. Recognizing the importance of contextualizing 
our findings across different cultural and regional settings, we 
also included country‐specific keywords such as ‘Sri Lanka’, 
‘India’, ‘Taiwan’, and ‘Morocco’. Given the research focus on 
food and health, we employed a range of domain‐specific 
terms such as ‘traditional foods’, ‘breakfast’, ‘diabetes’, 
‘glucose’, ‘blood sugar’, and ‘food technology’ to further refine 
our results. 

 
Table 1. Summary of the searching process 

Duration of the 
search 

Used research 
repositories 

Key words Type of 
Research 

works 

1st August 2024 
to  
30th September 
2024 

Scopus, Semantic 
Scholar, 
Sciencedirect, 
Scispace, 

Glycemic 
index, 
Glycemic 
index predict, 

Research, 
Thesis, 
Review 
Articles, Book   

IEEE Xplore,      
Digital Library, 
Google Scholar 

machine 
learning 
algorithm 
Glycemic 
Index, 

Chapters, 
Conference 
Materials, 
Reports 

 
 

Glycemic 
 

  
Index for 
machine 

 

  
learning 

 

 

We conducted our literature search across well‐
established academic repositories such as Scopus, 
ScienceDirect, IEEE Xplore Digital Library, Google Scholar, 
and Semantic Scholar, ensuring the credibility and diversity 
of our sources. Our review began in August 2024, 
concentrating primarily on research articles, review papers, 
book chapters, and conference proceedings, all published in 
English. The screening process involved two reviewers. As 
machine learning technologies gained momentum post‐1959, 
we focused our literature review on publications from 1960 
to the present, ensuring that we captured the full breadth of 
developments in this field. Summarized information is given 
in Table 1. The reviewed literature was further categorized 
based on its type, such as journal articles, conference 
proceedings, and other formats (Figure 2). 

 

 

The selected papers were categorized based on the key 
technologies used to derive findings related to the Glycemic 
Index. The study aims to highlight the significance of the 
Glycemic Index and assess its value in various research 
contexts. Our focus includes: 
• The diverse technologies applied in Glycemic Index 

research and how they have been utilized. 
• Standard datasets employed in the selected studies. 
• Future directions and research opportunities in research 

based on the prediction of the Glycemic Index. 
For the article search, the repositories listed in Table 1 were 
used. Well-known repositories such as PubMed, Web of 
Science, and the Cochrane Library were not included, as they 
often require institutional subscriptions to retrieve full-text 
papers. Initially, 100 papers were collected in total, and 
information such as the year, authors, and paper title was 
added to an Excel sheet. The search query applied was 
("Glycemic index" OR "Glycemic index prediction" OR 
("machine learning algorithm" AND "Glycemic index") OR 
"Glycemic index for machine learning"). This search query 
was utilized to retrieve studies focusing on machine learning 
algorithms for predicting or analyzing GI. Machine learning 
was used in the search string, as ML is a key area in AI for 
predictions and includes deep learning, XAI, and other related 
technologies.  

Figure 2. Proportion of papers by category 

Of these, 66 articles were directly selected from 
repositories, while an additional 34 were discovered by 
reviewing the identified articles or through works by the 
same authors. Fifteen articles were excluded due to 
duplication, being written in languages other than English 
without available translations, or being irrelevant to the 
study.  However, some papers were not related to Computer 
Science and only described food technology and GI-related 
content. Such papers were removed after screening. 
Furthermore, the remaining papers were categorized 
according to different technologies, and eight major 
technologies, such as food technology, statistical techniques, 
NLP, RL, image processing, deep learning, and machine 
learning, were identified in our survey. Figure 3 provides a 
detailed summary of the article selection process. 

3. Overview of key technologies and concepts 

         This section provides an overview of the 
fundamental technologies and concepts pertinent to our 
literature survey. The discussion will cover key areas 
including Glycemic Index, Deep Learning, Machine Learning, 
Food Technology, Image Processing, Reinforcement Learning, 
Statistical Techniques, and Natural Language Processing.  

 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

96 

 

 
Figure 3. Flow chart of the comprehensive review process based on 
the PRISMA 

Although some areas are closely related, they are 
discussed and categorized separately to ensure a clearer and 
more transparent review process. This approach allows for a 
more structured analysis, making it easier to understand how 
each technology relates to the Glycemic Index and how it has 
been applied in various contexts. 

3.1 Glycemic index (GI) 
    Foods and beverages provide the body with energy 

through carbohydrates, fats, proteins, and alcohol. Among 
these macronutrients, carbohydrates are the body’s preferred 
source of energy. The Glycemic Index is a system that ranks 
carbohydrates in various foods and drinks based on their 
effect on blood glucose levels. Specifically, the GI measures 
how much and how quickly a particular food raises blood 
sugar levels after it is consumed. This index typically ranges 
from 0 to 100, with pure glucose set as the reference point at 
a value of 100. GI values can be categorized into three ranges: 
• Low GI:55 or less 
• Medium GI:56 to 69 
• High GI:70 to 100 
Foods with a high GI value (greater than 70) are rapidly 
digested and absorbed, causing a rapid increase in blood 
glucose levels. On the other hand, foods with a low GI value 
(less than 55) are digested and absorbed more slowly, 
resulting in a slower and more gradual increase in blood 
glucose levels. Foods high in refined carbohydrates and sugar 
are digested more quickly and often have a high GI; whole 
foods high in protein, fat, or fiber typically have a low GI. 
Foods that contain no carbohydrates, such as meat, fish, 
poultry, nuts, seeds, herbs, spices, and oils, are not assigned a 
GI value [11]. The Glycemic Index is calculated by measuring 
the blood glucose response of a group of people after they 
consume a specific food, typically using a standard amount of 
carbohydrate (usually 50 grams). The Area Under the Curve 
(AUC) for the blood glucose response over a two‐hour period 
is measured, and the GI is determined by comparing the AUC 
of the test food to that of the reference food (either glucose or 
white bread, which are used as reference foods). The formula 
for calculating the GI is as follows: 

𝐺𝐼 =
(𝐴𝑈𝐶 𝑜𝑓 𝑡𝑒𝑠𝑡 𝑓𝑜𝑜𝑑)× 100

𝐴𝑈𝐶 𝑜𝑓 𝑅𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒 𝑓𝑜𝑜𝑑
                                     (1) 

To calculate the GI of a meal, one must know the GI values 
of the individual components. For example, if a meal consists 
of one cup of cooked brown rice (GI = 50) and one serving of 
grilled salmon (GI = 25), the total GI for the meal would be 75, 
which reflects the combined effect of these foods on blood 
sugar levels. Other factors that affect the GI of a food include 
its ripeness, cooking method, type of sugar it contains, and the 
amount of processing it has undergone. Understanding the 
Glycemic Index and the factors that influence it can help 
individuals make informed dietary choices, particularly those 
managing conditions such as diabetes [12]. 

As GI appeared as one of the promising approaches to 
identify the levels of carbohydrates, there is extensive 
literature combining GI with various perspectives, including 
food technology, nutrition science, and medical research. 
Although the primary objective of this survey is to explore the 
technological perspective of GI, it is important to 
acknowledge relevant past works in the food technology 
domain. These studies provide valuable insights into the 
nutritional impact and health benefits of foods with varying 
GI levels, as well as methods for modifying GI through food 
processing techniques. By including these works in the GI 
introduction section, we aim to provide a comprehensive 
background that contextualizes the technological 
applications we focus on, even though the main goal of this 
survey is not to delve deeply into food processing or 
nutritional studies. This approach helps to emphasize the 
interdisciplinary nature of GI research, while still keeping our 
focus on technological advancements and innovations in GI 
prediction and analysis. 

2024 ‐ 2015                [11],[17],[19],[20],[21],[22],[23],  
        [24],[25],[26],[27], [28],[29],[30], 
   [31],[32],[33],[34],[35],[36],[37], 
   [38],[39],[40],[41] 
 
2014 ‐ 2005 [15],[16],[42],[43],[44] 
 
2004 ‐ 1995 [14] 
  
1994 ‐ 1985 
 
1984 ‐ 1980 [13] 
 

Figure 4. Timeline of research where GI was used in Food Technology 

Early such work, mostly focused on calculating the 
Glycemic Index, by using food with different carbohydrate 
levels, and measuring blood glucose levels [13-15]. All three 
studies aim to determine the glycemic impact of foods. The 
first two studies follow the standard approach of measuring 
postprandial blood glucose response in humans, while the 
third study attempts to predict GI using a laboratory-based 
method. Similar studies, such as [16-19], highlight various 
approaches to measuring and applying GI, including 
standardization of measurement techniques, regional 
adaptations, and predictive modeling. One critical aspect here 
to address is the variability in methodologies, which can affect 
the reliability and comparability of GI values across studies. 
They may lead to inconsistencies in the results. Readers 
interested in exploring further can refer to Figure 4, which 
provides an overview of various GI-based approaches applied 
in food science research. 

 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

97 

 

3.2 Machine learning (ML) 
Throughout history, humans have continually developed 

tools to simplify tasks and meet various needs. The invention 
of machines was a major leap forward, revolutionizing areas 
such as transportation, industry, and computing. One of the 
most significant advancements in recent years is machine 
learning, a technology that has further extended the 
capabilities of these machines. Machine Learning focuses on 
enabling machines to handle data more effectively. In many 
cases, large datasets are too complex for humans to easily 
interpret. ML algorithms address this by identifying patterns 
and extracting meaningful insights from the data. As the 
availability of vast datasets continues to grow, so too has the 
demand for machine learning, with industries applying it to 
uncover valuable information. Unlike traditional 
programming, where explicit instructions are given, ML 
allows machines to learn from data and make decisions based 
on it. This shift has prompted researchers and engineers to 
develop approaches that allow machines to learn 
autonomously, without needing detailed programming for 
every task [45]. 

It’s important to recognize that machine learning is not 
just about managing data; it is also a crucial part of artificial 
intelligence. As a subset of AI, machine learning allows 
systems to discover hidden patterns within datasets, enabling 
them to make predictions about new data. This ability to 
generalize from previous experiences is essential for creating 
systems that can adapt to changing environments. For a 
system to be considered intelligent, especially in dynamic and 
unpredictable conditions, it must be able to learn and evolve. 
If a system can adapt to changes on its own, the designer does 
not need to foresee and program solutions for every possible 
scenario. This adaptability is one of the key strengths of 
machine learning. Machine Learning systems can be 
categorized based on various criteria. These include: 
• How they are trained (e.g., supervised, unsupervised, semi‐

supervised, self‐supervised) 
• Whether they can learn continuously in real time (online 

learning) or process data in batches (batch learning) 
• Whether they compare new data points to known data or 

build predictive models by detecting patterns (instance‐
based versus model‐based learning) [46]. 

These categories reflect the diversity of machine learning 
approaches, each tailored to address different types of 
problems and data environments. 

3.3 Reinforcement learning (RL) 
Reinforcement Learning is a type of ML in which an agent 

learns to make decisions by interacting with an environment 
and receiving feedback in the form of rewards or penalties. 
The agent’s goal is to maximize cumulative rewards by 
choosing actions that lead to favorable outcomes. Unlike 
supervised learning, where models are trained on labeled 
datasets, Reinforcement Learning relies on trial and error, 
allowing the agent to explore and exploit the environment to 
improve its strategy over time. The core components of RL 
include: 
• Agent: The learner or decision‐maker (Single/Multi) [47]. 
• Environment: The setting in which the agent operates. 
• Action: Choices made by the agent to interact with the 

environment. 
• State: The current situation or status of the environment. 
• Reward: The feedback the agent receives after taking an 

action 
At each step, the agent observes the current state of the 
environment, takes an action, and receives a reward based on 

the outcome. This process helps the agent learn a policy — a 
mapping from states to actions that maximizes long‐term 
rewards. Reinforcement Learning is widely used in various 
applications such as robotics, game AI, autonomous vehicles, 
and resource management. The Deep Reinforcement 
Learning (DRL) approach, which combines Deep Learning 
with RL, has significantly advanced the field by enabling 
agents to handle high‐dimensional, complex environments 
like images and continuous spaces. The challenge in RL lies in 
balancing exploration (trying new actions) and exploitation 
(choosing actions known to yield high rewards), ensuring that 
the agent learns an optimal strategy efficiently. 

3.4 Deep learning (DL) 
Deep Learning is a subset of ML that mimics the 

functioning of the human brain in processing data and 
creating patterns for decision‐making. It uses neural 
networks with multiple layers to model complex patterns and 
relationships in large datasets. DL has enabled remarkable 
advancements in areas such as Computer Vision, Natural 
Language Processing, and Speech Recognition. This 
revolutionary approach to machine learning has the potential 
to reshape various industries, including healthcare, where it 
is poised to drive significant improvements in medical 
imaging, disease diagnosis, and drug discovery [48-50]. The 
distinguishing feature of deep learning is its use of multiple 
layers of these artificial neurons, often referred to as “deep 
neural networks”. This depth enables the system to 
automatically extract features from raw data without the 
need for manual intervention or feature engineering. As a 
result, deep learning excels in tasks such as image and speech 
recognition, natural language processing, and even complex 
game strategies. DL has shown remarkable success in various 
applications, including self‐driving cars, medical diagnostics, 
and predictive analytics [51,52]. The availability of large 
datasets, along with significant advancements in 
computational power (especially through Graphics 
Processing Units (GPUs) and cloud computing), has 
contributed to the rapid development and adoption of deep 
learning techniques. Despite its successes, DL has challenges, 
such as the need for vast amounts of labeled data and high 
computational resources. Additionally, the models often act 
as “black boxes”, making their decision‐making process 
difficult to interpret. Nonetheless, the field of deep learning 
continues to evolve, pushing the boundaries of what 
machines can achieve in terms of intelligence and automation. 

3.5 Image processing (IP) 
Image Processing is a technique used to perform various 

operations on images to enhance their quality or extract 
meaningful information. It is a valuable tool for analyzing and 
transforming images, making them more suitable for specific 
applications or interpretations. Whether the goal is to 
improve visual quality, recover lost or degraded information, 
or extract critical details, image processing plays a vital role 
in fields such as computer vision, medical imaging, satellite 
imagery, and photography. At its core, image processing relies 
on computational algorithms that analyze the pixel data in 
images and apply a series of manipulations to achieve the 
desired outcome. These algorithms can be designed to 
address different aspects of an image, such as enhancing 
colors, sharpening details, reducing noise, or highlighting 
specific features. The complexity of these operations can 
range from simple tasks, such as adjusting brightness and 
contrast, to advanced techniques, like edge detection, object 
recognition, and image segmentation. There are two main 
types of image processing: 

https://docs.google.com/document/d/14wrzRJbMmXrNMvtgEGZciuAlQ2bP6c8u/edit#heading=h.1r7ewdq61d1s


NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

98 

 

• Analog Image Processing 
• Digital Image Processing 
Analog Image Processing involves handling images in a 
continuous signal form (e.g., photographs or X‐ray images) 
and is often used in traditional photography or medical 
imaging. Digital Image Processing involves converting images 
into a digital format and then processing them using 
computers. This type of image processing has widespread 
applications in fields like computer vision, medical imaging, 
remote sensing, facial recognition, and more [53]. Key tasks 
in Digital Image Processing include image enhancement, 
which focuses on improving the visual quality of an image, 
such as sharpening or adjusting contrast, and image 
restoration, which aims to remove noise or distortions to 
recover the original image. Image segmentation involves 
dividing an image into meaningful parts, such as identifying 
objects within the image, while image compression reduces 
the file size of an image, preserving its quality. Feature 
extraction is another crucial task, where key patterns or 
features in an image are identified for further analysis, often 
used in computer vision and machine learning. Common 
techniques in digital image processing include filtering, edge 
detection, histogram equalization, and Fourier transforms. 

3.6 Natural language processing (NLP) 
 Natural Language Processing is a field of artificial 

intelligence that focuses on the interaction between 
computers and humans through natural language. The goal of 
NLP is to enable computers to understand, interpret, and 
respond to human language in a valuable way. This field 
involves several tasks, including: 
• Text processing: This includes tokenization, stemming, 

lemmatization, and part‐of‐speech tagging to prepare text 
for analysis. 

• Sentiment analysis: Determining the emotional tone 
behind a series of words, used in applications like customer 
feedback analysis. 

• Named entity recognition (NER): Identifying and 
classifying key entities in text (e.g., names of people, 
organizations, locations). 

• Machine translation: Translating text from one language to 
another, as seen in tools like Google Translate. 

• Speech recognition: Converting spoken language into text, 
used in virtual assistants like Siri and Alexa. 

• Text generation: Creating coherent and contextually 
relevant text, such as chatbots or story generation. 

• Question answering: Developing systems that can answer 
questions posed in natural language, often used in 
customer support and search engines. 

Machine learning techniques are applied to textual data 
similarly to how they are utilized in other forms of data, 
including images, speech, and structured datasets. Supervised 
machine learning techniques, such as classification and 
regression methods, play a significant role in various NLP 
tasks. For instance, an NLP classification task might involve 
categorizing news articles into specific topics, such as sports 
or politics. Conversely, regression techniques can predict 
numeric values, such as estimating the price of a stock based 
on discussions in social media. Additionally, unsupervised 
clustering algorithms can be employed to group together 
similar text documents. Any machine learning approach for 
NLP, whether supervised or unsupervised, can be 
characterized by three common steps: extracting features 
from text, utilizing the feature representation to learn a 
model, and evaluating and refining the model [54]. 

 

3.7 Continuous glucose monitoring (CGM) 
Continuous Glucose Monitoring is a technology used to 

track glucose levels in real‐time throughout the day and night. 
It involves a small, wearable sensor inserted under the skin, 
typically on the abdomen or arm, which measures interstitial 
glucose levels at regular intervals. These readings are 
transmitted to a receiver or smartphone, providing users with 
a continuous stream of glucose data. This technology not only 
tracks glucose trends but also provides alerts for 
hypoglycemia or hyperglycemia, allowing proactive 
management of diabetes. CGM is widely used in diabetes 
management, particularly for individuals with type 1 and type 
2 diabetes, to improve glycemic control and reduce the risk of 
complications. 

3.8 Explainable AI (XAI) 
Explainable AI refers to artificial intelligence systems 

designed in a way that their decisions, predictions, and 
behaviors can be understood and interpreted by humans. The 
goal of XAI is to make AI more transparent, trustworthy, and 
accountable, especially in critical applications such as 
healthcare, finance, and autonomous systems [55]. Here’s a 
breakdown of the concept: Key Aspects of Explainable AI: 
• Transparency: The AI model provides insights into how it 

processes input data to produce its output. This might 
involve revealing the structure of the model, the logic 
behind decision‐making, or the importance of features in a 
prediction. 

• Interpretability: The results or decisions made by the AI are 
presented in a way that humans can understand. For 
instance, instead of presenting a decision as a ”black box” 
output, the AI explains why a specific choice or prediction 
was made. 

• Accountability: XAI systems allow developers, users, and 
regulators to scrutinize and validate the AI’s decisions, 
ensuring ethical and fair outcomes. 

• Trustworthiness: By making AI systems understandable, 
XAI builds confidence in their use, especially in high‐stakes 
scenarios where decisions impact lives. 

Traditional AI models, particularly those based on deep 
learning, often operate as “black boxes,” meaning their 
internal workings are complex and not easily interpretable. 
XAI addresses this limitation by providing insights into how 
and why an AI system arrives at specific outcomes, enabling 
users to trust and validate the model’s predictions. By 
fostering transparency, XAI enhances collaboration between 
humans and AI while reducing biases and errors in AI 
applications. Major Techniques in Explainable AI: 
• Intrinsic interpretability: Some models, like linear 

regression or decision trees, are inherently interpretable 
because their structure is simple and their outputs are easy 
to trace back to inputs. 

• Post‐hoc explanations: For complex models like deep 
neural networks, techniques are applied after the model 
has made predictions to explain the output. Common 
methods include: 

• SHAP: Quantifies the contribution of each feature to a 
prediction. 

• LIME (Local Interpretable Model‐agnostic Explanations): 
Builds interpretable models around individual predictions. 

• Feature importance analysis: Highlights which input 
features were most influential in a decision. 

• Visualization tools: For example, heatmaps in computer 
vision models show which parts of an image influenced a 
decision. 

 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

99 

 

SHAP, which stands for Shapley Additive Explanations, is 
an interpretability method grounded in Shapley values and 
was introduced by Lundberg and Lee. This approach has 
become widely adopted in machine learning to explain model 
outputs by quantifying the contribution of each feature to the 
final prediction, making complex models more transparent 
and interpretable. SHAP introduces two key innovations: (1) 
the identification of a novel class of additive feature 
importance measures and (2) theoretical results 
demonstrating the existence of a unique solution within this 
class that satisfies a set of desirable properties, such as local 
accuracy, consistency, and additivity [56]. These properties 
ensure that the feature attributions are both fair and reliable. 
This framework unifies six existing methods under a common 
theoretical foundation, offering a more robust and coherent 
approach to feature importance. Notably, it addresses 
shortcomings in several recent methods within this class that 
fail to satisfy the proposed desirable properties. SHAP’s 
interpretability extends beyond theoretical rigor, providing 
practical tools like visualization plots that enhance 
understanding of how features influence individual 
predictions and overall model behavior, thereby empowering 
users in high‐stakes domains like healthcare, finance, and law. 

 
3.9 Analysis of Variance (ANOVA) 

ANOVA is a statistical method used to determine 
whether there are significant differences between the means 
of three or more unrelated groups. Developed by Ronald 
Fisher, ANOVA extends the capabilities of the t‐test, which is 
limited to comparing only two groups. The primary function 
of ANOVA is to analyze how different categorical independent 
variables influence a continuous dependent variable by 
partitioning the total variance observed into components 
attributable to different sources. 

ANOVA operates under several assumptions: the 
samples must be independent, the dependent variable should 
be normally distributed, and the variances among the groups 
should be approximately equal (homogeneity of variance). 
The test statistic for ANOVA is the F‐value, calculated as the 
ratio of variance explained by the treatment (between‐group 
variance) to the variance due to random chance (within‐
group variance). A significant F‐value indicates that at least 
one group mean differs from the others, although it does not 
specify which means are different; post‐hoc tests are required 
for that purpose. 

There are various forms of ANOVA, including one‐way 
ANOVA, which examines a single independent variable with 
multiple levels, and two‐way ANOVA, which assesses the 
impact of two independent variables and their interaction on 
a dependent variable. This flexibility makes ANOVA a 
powerful tool for researchers looking to understand complex 
relationships in their data [57]. 

3.10 Tukey’s Q method 
The Tukey’s Q method, also known as the Tukey HSD 

(Honestly Significant Difference) test, is a statistical tool used 
to compare the means of different groups after conducting a 
one‐way ANOVA. It helps identify specific group differences 
when ANOVA indicates significant variance among groups but 
does not specify which groups differ. The Tukey HSD test 
calculates a statistic known as ‘q’, which is then compared to 
critical values from the Studentized range distribution. If the 
calculated ‘q’ exceeds the critical value, it indicates a 
significant difference between the group means. This method 
is particularly useful because it controls the experiment‐wise 
error rate, reducing the likelihood of Type I errors that can 

occur when conducting multiple t‐tests. By focusing on the 
largest pairwise differences in means, Tukey’s HSD provides 
a conservative approach to identifying significant differences 
while maintaining statistical rigor. Researchers often rely on 
statistical software to perform these calculations due to their 
complexity, but understanding the underlying steps, such as 
calculating overall and group means, sum of squares, and 
mean squares, is crucial for interpreting results accurately. 
Overall, Tukey’s Q method serves as an effective post‐hoc 
analysis tool in research studies where multiple group 
comparisons are necessary [58]. 

3.11 T‐test 
The t‐test is a statistical hypothesis test used to 

determine whether there is a significant difference between 
the means of two groups or between a sample mean and a 
known population mean. It is particularly useful when dealing 
with small sample sizes (typically n ≤ 30) and when the 
population standard deviation is unknown. There are three 
main types of t‐tests: the one‐sample t‐test, which compares 
a sample mean to a known value; the independent t‐test, 
which assesses the means of two independent groups; and the 
paired t‐test, which evaluates means from the same group at 
different times or under different conditions. The t‐test 
calculates a t‐value based on the difference between group 
means and their variability, which is then compared to critical 
values from the t‐distribution to assess statistical significance. 
This method helps researchers understand whether observed 
differences are likely due to chance or reflect true differences 
in the populations being studied [59]. 

4. Harnessing technology in Glycemic index research: 

innovations and insights 

Here, we focus on the main objective of this study: to 
enlighten the reader on how technology can be effectively 
utilized to tackle various challenges associated with the 
Glycemic Index. For convenience, we focus on specific 
technological aspects one at a time and discuss studies that 
have utilized them, either fully or partially, to address 
challenges related to the Glycemic Index, as illustrated in 
Figure 5, which presents a timeline of research where GI has 
been used with different AI-based technologies. 

4.1 Role of machine learning in research related to the 
Glycemic index 
Machine learning is a branch of AI and computer science 

that focuses on using data and algorithms to enable AI to 
imitate the way that humans learn, progressively improving 
its accuracy. Overall, ML is used to make decisions based on 
data. By modeling the algorithms on the basis of historical 
data, they find the patterns and relationships that are hard for 
humans to detect. These patterns are now further used for 
future reference to predict solutions to unseen problems in 
different domains. Biology and Food technology are some of 
the key domains that have used ML. Numerous studies have 
focused on the GI, exploring ML techniques to achieve diverse 
objectives. Given the increasing prominence of ML in GI‐
related research over the years (Figure 6), it is worth 
highlighting this category as a central focus of the discussion. 

The earliest record in our repository originates from 
2017: Glycaemic Index Prediction: a Pilot Study of Data 
Linkage Challenges and the Application of Machine Learning 
[63]. They present a ML‐based model that predicts the GI of 
foods based on the biochemical properties. They employed a 
multiple regression model, which bases its prediction on a 
weighted linear combination of the independent input 
variables. These variables include: (1) water (% of mass), (2) 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

100 

 

energy (kJ per 100g), (3) protein (% of mass), (4) total 
carbohydrates (% of mass), (5) sugars (% of mass), (6) fiber 
(% of mass), and (7) lipids (% of mass). They used GI as a 
target variable. In addition, the standard five‐fold cross-
validation methodology has been used for training and 
testing. Furthermore, they highlighted the need for the 
adoption of a common standard for recording different types 
of information on foods so that this information can be cross‐
linked automatically and without ambiguity. 

 
2024 ‐ 2015 [11],[17],[18], [19],[21],[22],[23], 

   [24],[25],[26], [27], [28], [29],[30], 

   [31],[32],[33],[34],[35],[36],[37], 

   [38],[39],[40],[41],[60],[61],[62],  

   [63],[64],[65],[66], [67],[68],[69], 

       [70],[71],[72],[73],[74],[75],[76], 

   [77],[78], [79],[80],[81],[82],[83] 

   [84],[85],[86],[87],[88],[89],[90], 

   [91],[92],[93],[94],[95],[96],[97], 

   [98],[99],[100],[101],[102]  

 

2014 ‐ 2005 [15],[16],[42],[43],[44],[103],[104] 

   [105],[106],[107] 

 

2004 ‐ 1995 [14], [108] 

 

1994 ‐ 1985    [109],[110]  

 

1984 ‐ 1980 [13] 

Figure 5. Timeline of research where GI were used with different AI-
based technologies  

 

Figure 6. No. of papers by year (overall and ML category) 

Colorectal Cancer (CRC) is recognized as the most 
preventable cancer worldwide. The GI has been used to 
assess healthy eating in association with CRC. The 
researchers explored predictors of the Healthy Eating Index 
(HEI) and GI in multi‐ethnic CRC families. In this study, GI 
served as one of the key measures of diet quality, helping to 

realize its role in managing CRC risk. Predicting GI and HEI is 
a major challenge in the real world. In this study, they 
employed machine learning techniques for validating and 
predicting HEI and GI. The validation procedures included the 
use of ensemble methods and generalized regression models, 
Elastic Net with Akaike’s Information Criterion with 
correction (AICc), and Leave‐One‐Out cross-validation 
methods. Generalized Regression (GR) models were 
employed with Elastic Net and validation methods (AICc and 
Leave‐One‐Out cross-validation) to minimize over‐fitting and 
to optimize prediction models for both HEI and GI. AICc 
validation and LOO cross‐validation methods are effective 
methods for small data sets. Results obtained revealed that 
further studies with larger datasets and diverse samples are 
needed to emphasize findings in diverse groups [66].  

In the same direction, another study aimed to validate 
predictors of healthy eating metrics: HEI, GI, and GL across 
various modern diets. The researchers examined daily 
dietary data from 131 diets classified into four primary 
groups (liquids, convenience foods, ethnic diets, and 
smoothies) to assess the impact of various diets on GI, GL, and 
HEI scores. Logistic Regression (LR) was used as a baseline 
model for initial predictions in this study. Also, Elastic Net 
Generalized Regression was applied for improved accuracy 
and to handle complex data with multiple predictors, and 
Elastic Net combines ridge and lasso regression, employed to 
minimize over‐fitting while allowing selection of relevant 
predictors [31].  In another work, Partial Least Squares (PLS) 
regression was applied to predict the GI and amylose content 
from Near‐Infrared (NIR) Spectroscopy to analyze rice 
characteristics spectral data. Near‐Infrared Spectroscopy 
(NIRS) is a primary technology that is used to collect spectral 
data from rice varieties in the wavelength range of 740–
1070nm. The model mainly used NIR as features to predict 
the GI and amylose content. Random Forest (RF) was 
employed for rice varieties classification, while the Principal 
Component Analysis (PCA) was used for dimensionality 
reduction to enhance the classification performance. Also 
Linear Discriminant Analysis (LDA) for classifying rice 
samples according to parboiling treatments was employed. 
This research also explored the use of a portable NIR sensor 
for real‐time, on‐site evaluation, creating predictive models 
for identifying rice varieties and estimating amylose content 
[77]. 

While the glycemic index may not be one of the most 
critical factors, it serves an indirect yet important role in 
certain situations. Intensive insulin treatment is a standard of 
care for tight glycemic control in people with diabetes to 
prevent or delay long‐term complications of diabetes 
mellitus. However, insulin therapy may trigger lethal 
hypoglycemia, and these results show that a number of 
subjects are prevented by this risk factor from attaining and 
sustaining near normoglycemia. The forecasting of 
postprandial hypoglycemia is considered to improve the CGM 
technology for persons with diabetes using insulin. The GI can 
also change the Rate of Glucose Increase (RIG), which is a 
predictor of hypoglycemia. The researchers did not use 
specific GI values, but used glucose profiles characteristic of 
high GI foods to help simulate hypoglycemia dangers. This 
study employed four machine learning models to predict 
hypoglycemia, including RF, Support Vector Machine (SVM) 
(with both linear and radial basis functions), K‐Nearest 
Neighbor (KNN), and Logistic Regression. Among the four 
models, the Random Forest was the best with an average of 
AUC 0.966 and was quite good at predictive ability. The 
researchers in this study plan to explore evaluation of their 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

101 

 

algorithm on a prospective patient population to clearly 
establish the clinical use of this system [71]. 

Another study pursued the usage of machine learning 
approaches for estimating short‐term blood glucose levels of 
Type 1 Diabetes (T1D) patients. T1D is an autoimmune 
disease in which the pancreas releases little or no insulin. The 
traditional method involves having patients administer 
insulin shots to themselves on a number of occasions each 
day. In this work, the focus was geared more towards proper 
identification of suitable ML models depending on glycemic 
status (hypoglycemia, normoglycemia, and hyperglycemia). 
This study also addressed the challenge of imbalanced data, 
which occurs when T1D patients spend the majority of their 
time in the normoglycemic range. GI affects how the study 
encompasses the dynamics of glucose related to food intake; 
it has an indirect effect on how blood glucose is managed and 
anticipated. Ten different machine learning and deep learning 
algorithms were used for training regression models, 
including Linear Support Vector Regression (SVR), Lasso 
Regression, Decision Trees, Random Forest, KNN, Multilayer 
Perceptron (MLP), and Gradient Boosting. The prediction 
models used 24 consecutive CGM sensor measurements 
obtained every five minutes over a 120-minute period, with 
the blood glucose level set 30 minutes after the last reading. 
Overall, the work provided a detailed overview of ML 
strategies for blood glucose prediction, highlighting the 
necessity of tailored models and oversampling techniques 
when dealing with imbalanced glycemic data [73].  

Similarly reference [75] aimed at establishing the 
feasibility of estimating blood glucose levels of  T1DM 
patients through constrained platforms like smartphones and 
Raspberry Pi. As the research objective, the real‐time glucose 
prediction from aggregated data streams was performed with 
the help of the Machine Learning models implemented on the 
local devices excluding the cloud computing that can provide 
predictions even if there is no internet connection. Data from 
CGM was used to develop univariate models in which 
forecasts are based on preceding glycemic values. Despite the 
fact that the model did not use GI values directly it relied on 
the CGM records which captured patterns of effects of high‐GI 
foods. The model was able to use the observed glucose 
changes and calculate the future glucose values without an 
accurate GI of all foods in the meal. This study employed ML 
models such as RF, SVM, and Autoregressive Integrated 
Moving Average (ARIMA). SVMs also perform computations 
efficiently on devices with restricted computational 
capabilities, like Smartphones and Raspberry Pi, and this was 
most evident with tiny sliding window computations. 
Heading in the same way, another research presented an 
ensemble machine learning approach to detect unannounced 
meals (UAM) in type 1 diabetes patients. Maintenance of 
postprandial glucose level is another considerable barrier in 
the management of T1D. CGM and hybrid Automated Insulin 
Delivery (AID) systems depend on patients to alert and 
predict carbohydrate (CHO) intake, which is frequently 
ignored, especially in adolescents. This study identified that 
missing meal announcements increase time to insulin 
administration, increase postprandial blood glucose 
variability, and reduce overall glycemic control. Here also, GI 
is indirectly used for glycemic control by aiming to detect 
UAM and improve glucose management. The ensemble model 
was built combining the predictions of three ML models: 
Artificial Neural Network (ANN), RF, and Logistic Regression. 
A total of 14 features were used, of which 12 are based on 
CGM readings and the remaining two are based on insulin 
data [97]. 

Maintaining glycemic control in children with type 1 
diabetes is a challenging task in clinical practice. A study has 
focused on predicting glycemic control (measured by 
glycated hemoglobin levels (A1C)) in children with Type 1 
Diabetes using machine learning algorithms. Binary Logistic 
Regression was applied to predict the probability of poor 
glycemic control, identifying significant predictors such as 
A1C at onset and ketoacidosis episodes. GI was not directly 
used in this study. In the model, the initial A1C level was an 
essential covariate because it captures historical glycemia 
that could be influenced by GI. High A1C could indirectly 
capture patterns associated with frequent intake of high‐GI 
foods if they led to sustained high glucose levels over time. 
This study used 15 features, including demographic and 
socioeconomic Factors like family income, living 
environment, maternal and paternal education. Overall, the 
indirect impact of GI on long‐term glycemia regulation could 
be conferred in baseline A1C and lipid profile captured by the 
model [90]. 

The main types of diabetes are Type 1, which comprises 
5‐10% of diabetes patients. According to the International 
Diabetes Federation (IDF, 2017), Type 1 diabetes is caused by 
an autoimmune reaction in which the body’s immune system 
attacks the insulin‐producing beta cells of the pancreas and 
causes the body to produce very little or no insulin; hence a 
diabetes patient is required to ad‐ minister insulin on daily 
basis to maintain the recommended target blood glucose 
level. Type 2, which was formerly well known as non‐insulin 
dependent, and which comprises 90‐95% of diabetes 
patients, is caused by the human body’s inability to fully 
respond to insulin (IDF). Focusing on Type 2 diabetes 
patients, a study aimed to develop a personalized food 
recommendation system. GI was the core metric to classify 
foods into high, medium, and low categories, helping the 
system to recommend foods with a lower likelihood of 
causing blood sugar spikes. They considered GI, GL, and 
carbohydrate content as features of the study. These features 
focused on the glycemic impact and carbohydrate content of 
food items to classify foods into categories. Naive Bayes 
Classifier was selected as the primary machine learning 
model to recognize patterns in glycemic response. Foods with 
known GI values were sourced from an international 
Glycemic Index database, which supports the model in 
recommending foods based on how they are likely to affect 
blood glucose [85].  

In tandem, another study tried to create a model 
predicting the GI of fruits. They employed a combination of 
DL and ML methods to predict the GI of fruits. The output of 
the consequent module can identify three fruits, including 
Apples, Bananas, and Oranges, but the GI is predicted only for 
bananas. They used bananas for GI prediction because the GI 
of bananas varies a large amount according to ripeness cycles 
and is a simple food for testing. The researchers used 
Glycemic Load (GL) to assess the overall Glycemic Index of the 
fruit. A Convolutional Neural Network (CNN) was used to 
determine the type of food. Then, a simple binarization model 
was used to characterize the measurements of the fruit 
length. They used the THRESH_BINARY function from the 
OpenCV library to binarize the image of the fruit. Two linear 
regression models were then applied to the prediction length 
to derive the GL and Carbohydrate content of the fruit, 
respectively. The pretrained machine learning models were 
invoked based on the ripeness of the fruit to predict the GL 
and the carbohydrate content of the fruit. Both linear 
regression models took length as an input parameter. Once 
the carbohydrate content and GL values were in, the data 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

102 

 

were plugged into the glycemic index formula to determine 
the Glycemic Index. Clearly, this research employed CNN for 
fruit classification and linear regression for GL prediction to 
help Pre‐Diabetes patients select healthy fruits according to 
their glycemic response [30]. 

In some research, the GI is used as a feature to evaluate 
and optimize dietary impacts on blood glucose levels. One 
such study focused on developing a machine learning model 
to predict postprandial blood glucose responses in patients 
with Gestational Diabetes Mellitus (GDM). A gradient 
boosting algorithm was used, and they extracted data from 
various resources like mobile app diaries, CGM, and 
individual patient characteristics to improve blood glucose 
control. The GI was used as one of the features in this model. 
The model also used GL, which combines the GI and the 
quantity of carbohydrates consumed to estimate the impact 
of food on blood glucose levels. They used random grid search 
and cross‐validation for hyperparameter tuning. Values from 
the SHAP method were employed to understand the influence 
of different features on model prediction. This study used 
explainable AI methods through SHAP to evaluate the impact 
of features on the predictions [78].  

On the same line, reference [39] employed GI as a 
feature. The study focused on increasing the prediction 
accuracy of the Postprandial Glycemic Responses (PPGR) in 
women with Gestational Diabetes Mellitus by incorporating 
GI and GL. CGM & food diaries of pregnant women were used 
for the development. They compared models with and 
without GI/GL data to determine whether or not GI/GL data 
can improve the prediction of PPGR. The study was focused 
on finding the effect of GI/GL information to enhance the 
prediction of PPGR outcomes. This study used a total of 124 
participants (90 GDM & 34 controls) from the prospective 
multi-center GEM‐GDM (Genetic and Epigenetic Mechanisms 
of Developing Gestational Diabetes Mellitus) clinical trial. 
Each of 1,489 meal records was associated with glucose 
measurements. GI values were derived directly from the 
University of Sydney database (available until October 2023) 
[110] and matched to foods in the DiaCompanion app food 
database. ML methods such as linear regression and 
regularized regression (Lasso, Ridge, Elastic‐Net, LARS Lasso, 
Orthogonal Matching Pursuit) were used in the study. 
Another study addressed the need for accurate, automated 
dietary monitoring by analyzing the Post‐Prandial Glucose 
Response (PPGR) to predict meal macronutrient content. 
They also used GI as a key factor influencing PPGR, especially 
for carbohydrate‐heavy foods. But the model did not directly 
calculate GI values. Instead, it used CGM‐based PPGR data to 
infer macronutrient compositions of meals. The model 
implicitly considered the impact of carbohydrates (via PPGR 
patterns) in estimating these compositions, which are 
indirectly related to GI effects. They evaluated the sparse‐
coding approach against two baseline techniques: (1) ridge 
regression (RR), as a representative of regularization 
methods, and (2) a Nearest‐Neighbor classifier operating in a 
Linear Discriminant Analysis subspace (LDA‐kNN), as a 
representative of distance-based classifiers [82]. In another 
similar work, machine learning models were used to predict 
the progress of the glycemic values of six patients with 
diabetes. Eight different algorithms were compared, i.e., ANN 
with Multilayer Perceptron, Probabilistic Neural Network 
(PNN), Polynomial Regression, Gradient Boosted Trees 
Regression, Random Forest Regression, Simple Regression 
Tree, Tree Ensemble Regression, and Linear Regression. The 
algorithms were classified based on the ability to minimize 
four statistical errors, namely: Mean Absolute Error, Mean 

Squared Error, Root Mean Squared Error, and Mean Signed 
Difference. Direct use of GI is not presented. Instead, it aimed 
to predict overall glycemic status using historical glucose 
readings from patients [89].  

In reference [88], researchers used wearable device data 
to attempt to predict future glycemic control among adults 
with prediabetes. In this study, they have 16 features, 
including physical activities, heart rate, and sleep. They aimed 
to predict longitudinal continuous changes in hemoglobin 
A1C and assess worsening, improvement of glycemic control 
among non‐diabetic and prediabetic adults using various 
features obtained from wearables. Directly calculated or 
predicted GI was not specifically measured in the study. 
Instead, changes in glycemic control were monitored using 
hemoglobin A1C levels (which reflect long-term blood 
glucose levels rather than short-term postprandial responses 
to foods). 

Diabetic Retinopathy (DR) is one of the major 
complications of diabetes. A recent study integrates ML 
models to predict the risk of diabetic retinopathy [99]. SHAP 
was established to increase the accuracy of risk prediction for 
diabetic retinopathy, explain the rationality of the findings 
from model prediction and improve the reliability of 
prediction results. The features that used in the model were 
extracted from a diabetes complication dataset. The CatBoost 
model was employed and optimized for the prediction task. 
The GI itself is not directly used as a feature in this study; 
instead, this study focused on glycemic measures like 
glycated hemoglobin (HbA1c) and fasting blood glucose 
(GLU_2H) as significant indicators in diabetic retinopathy risk 
assessment. GI measures how quickly carbohydrate-
containing foods raise blood glucose, which is particularly 
relevant for diabetes management. ML has yielded stunning 
success in predicting the importance of diabetes risk using 
health indicators and pattern analysis. 

“Diabetes Prediction Using Machine Learning 
Classification Algorithms” [111] reveals the effectiveness of 
several classification algorithms, including SVM, Extreme 
Gradient Boosting (XGB), Decision Trees (DT), and RF, in 
predicting diabetes. Implementing ML models in GI 
predictions would lead to identifying the effect of foods on 
blood glucose levels, thus helping to formulate dietary 
recommendations and support glucose management within 
both diabetic and pre‐diabetic populations. The glycemic 
variability metric is an additional measure available to the 
clinician as a potentially useful tool for estimating overall 
glycemia. Here, the researchers employed a new measure, 
Consensus Perceived Glycemic Variability (CPGV), for how 
much a patient’s blood glucose levels fluctuate, as evaluated 
by doctors. ML models were used to forecast blood glucose 
levels for 30 and 60 minutes in the future. This study focused 
on blood glucose variability and predicting blood glucose 
levels based on CGM data, which are different from the GI. 
Glycemic variability measurement and blood glucose 
prediction were modeled with 26 features. The CPGV metric 
was created using Linear Regression, whereas the future 
glucose levels were predicted using SVR and MLP [107]. 

“Application of Machine Learning Algorithms to Predict 
Uncontrolled Diabetes Using the All of Us Research Program 
Data” [91] used ML techniques to effectively predict 
uncontrolled diabetes using clinical markers such as serum 
electrolytes, body weight, and other physiological indicators. 
While the GI was not applied directly as a feature in this study, 
it demonstrated the use of ML as a potential tool for diabetes 
control using various predictors of glycemic status. This gives 
GI prediction an additional dimension, making it possible to 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

103 

 

adapt ML not only to the amounts of food we are consuming 
but also to provide a more holistic perspective on factors 
related to the regulation of blood glucose levels. This 
suggested the need for including GI as a supplemental dietary 
characteristic to improve personalized diabetes management 
in such models. To emphasize this, they employed techniques 
such as RF, Extreme Gradient Boosting (XGBoost), Logistic 
Regression, and Weighted Ensemble Model (WEM) that 
combines RF, XGBoost, and LR models. A similar study 
explored the relationship between noninvasive wearables 
and glycemic metrics and demonstrated the feasibility of 
using non-invasive wearables to estimate glycemic metrics, 
including hemoglobin A1c and glucose variability metrics 
[87]. The application was in real‐time for people with pre‐
diabetes or high‐normal glucose. ML approaches such as 
random forest models were used to estimate HbA1c levels. 
The study did not directly address the concept of GI but 
accentuated the potential for continuous, noninvasive 
monitoring of glycemic metrics. For this purpose, digital 
biomarkers were used to obtain physiological data, including 
skin temperature, electrodermal activity, heart rate, and 
accelerometry. The metrics correlate with glucose levels. 
Along with this, they provide insight into glucose variability 
without requiring traditional invasive measurements like 
blood samples. 

The maintenance of glycemia in range is one of the 
biggest challenges in the treatment of patients with diabetes. 
In a recent comparative study, the focus was to compare 
different diabetes management therapies, with an idea of the 
effectiveness of a machine-learning-trained closed‐loop 
artificial pancreas system. Diabetes Type 1 (DT1) patients' 
data were used in this study. They tried to measure 
improvements in glycemic control when switching from 
traditional therapies to the ML‐trained system. The Low 
Blood Glucose Index (LBGI) and High Blood Glucose Index 
(HBGI) were employed as blood glycemic indices that 
measure the likelihood of hypo and hyperglycemia events, 
respectively. Specific features used to predict glycemic 
control include time in range (TIR%), mean and median blood 
glucose levels, percentages of hypoglycemia and 
hyperglycemia, LBGI, HBGI, and glycated hemoglobin. The 
closed‐loop artificial pancreas algorithm was trained using 
machine learning techniques to optimize insulin dosage 
based on collected glucose data. Two ML regression 
techniques were tested by them in the R environment. This 
study concluded that “hybrid closed‐loop” artificial pancreas 
with control algorithm trained with machine learning 
technology provides very significant improvement in 
glycemia control compared to the multi‐daily injection (MDI), 
insulin pump without CGM, and sensor-assisted insulin pump 
therapies [76]. 

A study investigated the relationship between glycemic 
control, hyperhomocysteinemia, and microalbuminuria, 
which is an early marker of kidney and cardiovascular 
complications in diabetics [74]. They defined glycemic 
control by Fasting Blood Glucose (FBS) and glycosylated 
hemoglobin to assess their relations with microalbuminuria. 
The analysis for the association of urinary microalbumin with 
age, gender, HbA1c, FBS, and diabetic status was performed 
by using a multiple linear regression model. In another recent 
study, the GI is used as a conceptual foundation for 
understanding the glycemic impact of food. Estimation of the 
glycemic impact of cooking recipes using online 
crowdsourcing and machine learning is a novel approach. 
This study focused on glycemic impact, which refers to how a 
recipe affects blood sugar levels post‐consumption. The 

researchers considered the sugar‐to‐fiber (S/F) ratio as a 
proxy for the glycemic impact during the initial stages of 
recipe selection and modeling. Several ML models were 
developed, including Logistic Regression and LightGBM. 
Furthermore, the study experimented with various NLP 
techniques, such as bag‐of‐words (BoW) and word 
embeddings (e.g., Word2Vec, GloVe, FastText). They used 
both textual features and 20 nutrition features. Textual 
features included recipe titles, ingredients, and cooking 
directions. Nutrition features included carbohydrates, 
protein, fat, and dry weight. As limitations, they highlighted 
that the models trained on small datasets are prone to 
overfitting [72]. 

In addition, there were several review papers related to 
ML and GI. They were helpful in identifying the existing 
methods and the gaps related to the field. One of the recent 
studies has investigated the role of machine learning in 
nutrition science and diabetes management [102]. The article 
reviewed machine learning methods for screening Food 
Bioactive Compounds (FBCs) with bioactivities like 
antioxidant, anti‐inflammatory, antihypertensive, and 
hypoglycemic effects. It presents an ML model development 
process, covering data preparation, molecular 
representation, ML algorithm selection, and evaluation 
methods. SVM, RF, and KNN are commonly used for initial 
screening. And also CNN and Recurrent Neural Networks 
(RNN) are applied to complex data. This study further 
accentuates the importance of model interpretability. 
Techniques such as feature importance were used to 
understand the contribution of each molecular descriptor to 
the prediction of bioactivity. There was another synthesis 
that examined the role of AI and ML in outcomes to improve 
glucose control [94]. They fixated on predictive modeling 
development in the space of not only automated insulin 
delivery systems but also CGM. The study specifically worked 
on challenges faced in terms of data consistency, clinical 
accuracy, interpretability, and personalization. The study 
appears as a guide for ML practitioners on diabetes data, 
including best practices, feature engineering, standardizing 
datasets, and evaluating models. Overall, the findings of this 
review describe improvements due to ML in diabetes 
management. The paper also discussed the difficulties of 
applying ML and AI techniques, including the data processing 
inhomogeneity, metrics evaluation for models, and the usage 
of multiple data sources accounting for glycemic control 
interpatient variability. Going along the same direction as a 
literature analysis, another paper investigated the 
incorporation of GI into smartphone‐based food classification 
and nutritional estimation [95]. Most of the systems 
discussed in this review use computer vision to categorize 
foods and predict portion volumes to help facilitate dietary 
monitoring in diabetes management. Further, the paper 
reviewed possible future GI integration supportive 
technologies. ML and DL techniques have been thoroughly 
reviewed for food identification and volume estimation. CNNs 
such as AlexNet, VGG, ResNet, and EfficientNet are commonly 
used as baseline models to classify food images. Classification 
based on extracted features was performed using SVM and 
RF. According to these review studies, preliminary 
knowledge on ML has been integrated for food bioactivity 
screening, automated dietary assessment, and GI control. 
Research gaps still exist because no predictive modeling 
dedicated to the GI was found. Recently, machine learning 
approaches have been developed to estimate GI or inter-
individual glycemic responses than previous studies. These 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

104 

 

models are based on complex patterns of food composition 
variables in larger and more diverse datasets. 

Machine Learning technologies play a crucial role in 
research related to the GI. GI prediction, glycemic control, and 
how glycemic indices affect a diabetic patient were identified 
using ML algorithms. For such a task, identifying complex 
relationships between food composition, blood glucose levels, 
and health outcomes was taken into account. Regression 
models and classification models played a major role. 
Regression models such as multiple linear regression, elastic 
nets, and ridge regression were used to predict GI by 
analyzing food characteristics. These models provide a basis 
for calculating how different foods affect blood glucose levels. 
Classification algorithms, such as random forests, SVM, and 
logistic regression, enable food classification based on GI 
levels and have proven useful in creating personalized dietary 
recommendations. Time series models such as ARIMA and 
autoregressive neural networks are used in real‐time 
glycemic monitoring. By analyzing CGM data, it predicted 
blood glucose trends, which is especially helpful for diabetics 
who need dynamic management of blood sugar levels. 
Ensemble models, such as gradient boosting and random 
forest ensembles, improve predictive accuracy by combining 
multiple algorithms, often for applications such as identifying 
unreported foods or adjusting insulin in artificial pancreas 
systems. Finally, explainable AI models, using methods such 
as SHAP, provide insight into feature importance, allowing 
clinicians and researchers to understand which factors most 
influence glycemic outcomes. Together, these ML models 
form a robust framework for predicting, monitoring, and 
managing glycemic responses, making them central to the 
advancement of personalized glycemic control. Collectively, 
these approaches highlight how ML techniques can be applied 
not only to GI prediction but also to broader applications such 
as diabetes management, dietary assessment, and 
personalized nutrition by advancing computational nutrition 
science. 

4.2 Role of reinforcement learning (RL) in research 
related to the Glycemic index 
Reinforcement Learning is a machine learning paradigm. 

Agents learn to make decisions by interacting with an 
environment to maximize cumulative rewards. Unlike 
supervised learning, which is based on labeled data, RL 
involves trial‐and‐error exploration. This dynamic learning 
approach is particularly effective in problems that require 
sequential decision‐making, such as feature selection, control 
systems, and real‐time predictions. In our repository, the 
study “Impartial Feature Selection Using Multi‐Agent 
Reinforcement Learning for Adverse Glycemic Event 
Prediction” [20] represents a pioneering effort in applying RL 
to feature selection in the context of blood glucose prediction. 
They presented a model for predicting adverse glycemic 
events (normoglycemia, hypoglycemia, hyperglycemia) using 
CGM, Electronic Medical Record (EMR), Multi‐agent 
Reinforcement Learning (MARL), and Time2Vec (T2V). EMR 
data were used for feature selection. MARL employed optimal 
feature selection and selected optimal EMR features for better 
model performance. Although the study does not directly 
predict or calculate the GI of foods, it utilizes CGM‐derived 
blood glucose levels and EMR data to predict adverse 
glycemic events. MARL evaluated individual feature 
contributions and derived the optimal feature set by 
dynamically assigning rewards proportional to the 
performance change each feature contributed. It has been 
observed that, aside from this study, few significant efforts 

have been made to apply Reinforcement Learning (RL) to 
Glycemic Index research. This highlights an open area of 
exploration, presenting an opportunity to go deeper into the 
potential of RL in GI‐focused studies. 

4.3 Role of deep learning in research related to the 
Glycemic index 
Deep Learning is a specialized area of artificial 

intelligence that utilizes neural networks with multiple layers 
to analyze and interpret complex data patterns. By mimicking 
the way the human brain processes information, deep 
learning models can automatically extract features from raw 
data, making them highly effective for tasks such as image and 
speech recognition, natural language processing, and more. 
These models are trained on large datasets, adjusting their 
internal parameters to improve accuracy and performance. 
The rise of deep learning has been fueled by advancements in 
computational power and the availability of vast amounts of 
data, leading to significant breakthroughs in various fields, 
including healthcare, finance, and autonomous systems. One 
emerging area where deep learning is making a considerable 
impact is in health‐related research, specifically the 
prediction and recommendation of foods based on their GI, 
making it a critical tool for managing diabetes and other 
metabolic conditions. Through deep learning techniques, 
researchers can analyze food images, predict GI values, and 
recommend lower-GI alternatives to support personalized 
dietary plans. These applications combine computer vision 
with nutritional science, demonstrating deep learning’s 
potential to support better health outcomes through dietary 
management. 

Reference [44] is the earliest study in our survey, linking 
deep learning with GI prediction, specifically an artificial 
neural network, which is employed to predict the GI of foods. 
The process involves simulating human digestion, where 
samples undergo enzyme digestion, and their sugar content is 
analyzed using HPLC (High‐Performance Liquid 
Chromatography). The ANN takes the compositional data 
(such as protein, fat, dietary fiber, and sugar content) from the 
HPLC results and predicts the GI by learning from known GI 
values. The ANN model achieved a high correlation (r² = 0.93) 
between predicted and actual in vivo GI values, which shows 
that it could predict GI values closely matching those obtained 
from conventional human testing. Another way to predict GI 
is to analyze captured signals from chewing and swallowing, 
leading to our next research work, [67]. This proposes a 
method for managing diabetes by monitoring food intake 
behavior (chewing, swallowing, and saliva secretion) and its 
impact on blood glucose levels. The study uses a 
Microelectromechanical System (MEMS) acoustic sensor to 
capture signals from chewing and swallowing, analyzing 
these signals to predict and control postprandial GI. 
Convolutional neural networks are used for feature 
extraction from acoustic signals generated during chewing 
and swallowing, focusing on spatial and frequency patterns. 
Additionally, Deep Belief Networks (DBNs) are employed to 
further analyze non‐linear relationships in chewing signals, 
helping to generalize patterns and link them to blood glucose 
levels. When we talk more about health monitoring and 
predicting GI, the previous work [46] also focuses on 
predicting the GI of fruits, primarily focusing on bananas. The 
CNN model is applied for fruit recognition (apples, bananas, 
and oranges) and ripeness (raw‐green, ripe, overripe) 
detection. Ripeness is critical because the GI of bananas varies 
significantly with ripening. After that, the model uses image 
binarization to estimate banana length, exploring OpenCV’s 

https://docs.google.com/document/d/14wrzRJbMmXrNMvtgEGZciuAlQ2bP6c8u/edit#heading=h.1r7ewdq61d1s


NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

105 

 

Thresh_Binary and boundingRect functions. Length serves as 
an essential feature for GI and carbohydrate content 
prediction. The GI is derived from GL and carbohydrate 
content. The model assesses whether the fruit is safe to 
consume based on GI thresholds, providing dietary 
recommendations if necessary.  

Another fruit‐based study [93] uses deep learning 
techniques, specifically an improved Faster R‐CNN model 
with a Squeeze‐and‐Excitation (SE) attention module, to 
estimate the GL index of fruits, not the GI. The model identifies 
the fruit type (using R‐CNN) and estimates its volume (based 
on fruit size relative to a reference object (e.g., thumb) for 
accurate GL calculation), which are then used to calculate the 
GL based on existing GI values. The GI of each fruit type is a 
known input in this process. Instead, the model uses the 
fruit’s GI, volume, and carbohydrate content to compute the 
GL, providing a measure of how consuming that fruit might 
affect blood sugar levels. Just like fruits, bread is also explored 
as a research area for GI prediction through deep learning. 
[26] utilizes deep learning and computational intelligence 
methods to predict the digestion kinetics and GI (with the 
help of sample concentration of Euryale Ferox Seed Shell 
Extract (EFSSE), digestion time, and hydrolyzed starch 
concentration after digestion) of bread fortified with EFSSE. 
The Swarm Intelligence Supervised Neural Network (SISNN), 
specifically using Particle Swarm Optimization (PSO), 
simulates digestion kinetics more accurately than traditional 
mathematical modeling, aiding in predicting the glycemic 
response of fortified bread. The model demonstrates 
improved performance in predicting the GI of bread samples 
by integrating the optimization strengths of PSO with neural 
network modeling. 

Rather than predicting GI, reference [84] aimed to 
predict heart disease risk in diabetic patients using deep 
learning techniques. GI is used as part of the input data for 
predicting heart disease risk among diabetic patients, 
providing insights into how certain foods affect blood glucose 
levels. In this study, LSTM (Long Short‐Term Memory) was 
tested to determine its effectiveness in predicting heart 
disease based on diabetic patients’ data. Gated Recurrent Unit 
(GRU) is another RNN variant designed to handle sequential 
data, but with a simpler architecture compared to LSTM. GRU 
outperformed LSTM, providing better results in terms of 
accuracy and efficiency. It optimizes the learning rate through 
backpropagation, adjusting parameters to improve 
prediction accuracy. While the earlier works utilized deep 
learning to analyze food intake behaviors, the study [83] 
shifts focus toward comparing predictive models for 
continuous glucose monitoring. The comparison was carried 
out between ARIMA models for auto‐adaptive parameter 
tuning with statistical tests for real-time GI prediction and 
LSTM-based RNNs to capture long‐term dependencies, 
trained with backpropagation through time. This presents a 
novel method for parameter optimization in ARIMA and 
evaluates these models in a practical online learning scenario, 
with specific applications in health monitoring systems. 

Health monitoring alone is insufficient for maintaining 
good health; it is essential to consume appropriate food 
varieties and quantities to achieve a healthy lifestyle. In 
reference [25] GI is incorporated into a recommendation 
system that suggests healthier food alternatives for users 
with specific health conditions like diabetes. When a user 
inputs a food image, the model identifies the food item and 
retrieves its nutritional content, including GI. If the identified 
food has a GI over 55, the system recommends three similar 
foods with a lower GI, suitable for users who need to manage 

blood sugar levels. The study uses the InceptionV3 deep 
learning model, a CNN, for food image recognition, which 
achieved an accuracy of 75%. Another similar food image 
recognition application is used in the reference [11] aims to 
develop a Moroccan food dataset for food image recognition 
and nutritional analysis, specifically focusing on estimating 
the GI of various Moroccan dishes using deep learning 
techniques. The study employed transfer learning using pre‐
trained CNN models, specifically evaluating DenseNet, 
MobileNet, and EfficientNet. The GI and GL were calculated 
based on recognized food items, utilizing established GI 
databases and the carbohydrate content of the dishes. Foods 
were categorized into low, medium, and high GI and GL based 
on their values. 

In our collection related to DL, we found one review 
article [95] combining GI and deep learning techniques. The 
GI is used in these systems as a benchmark to guide dietary 
recommendations, particularly for diabetics. Once the system 
classifies a food item, it uses its estimated GI to assess 
potential blood glucose impact, and it can offer lower‐GI 
alternatives if needed. This approach aims to help diabetic 
patients manage post‐meal blood glucose levels by suggesting 
healthier food options. Here are some inputs used to predict 
GI and Nutritional Estimation in the previously mentioned 
studies done with DL and GI.  
• Food images: Captured by smartphone cameras and 

processed through CNNs for classification. 
• Volume estimation: Uses either single or multi‐view images 

to approximate portion sizes, which are critical for 
calculating nutrient intake. 

• Nutritional database: Contains data on each food item’s GI, 
carbohydrates, protein, and fats, enabling the model to 
provide personalized dietary advice. 

DL has become a transformative approach in GI prediction 
and dietary management by combining computer vision, 
signal processing, and recommendation systems. DL models 
such as CNNs, LSTMs, and GRUs can be used to understand the 
nonlinear behavior of food composition, eating behavior, and 
glycemic responses. DL-based applications span estimating GI 
using food images and detecting ripeness to provide 
personalized meal plans and continuous glucose monitoring. 
The integration of attention mechanisms, transfer learning, 
and optimization algorithms has further enhanced the 
accuracy of predictions. Summarizing the studies related to 
DL with GI, it can be stated that deep learning techniques can 
be effectively applied to predict the Glycemic Index of various 
foods, offering promising insights for more accurate and 
scalable GI estimation. By exploiting complex patterns within 
the data, deep learning models provide a robust approach 
that can be adapted to various types and characteristics of 
food, contributing significantly to advances in personalized 
nutrition and dietary recommendations. 

4.4 Role of image processing (IP) in research related to 
the Glycemic index 
Image processing has revolutionized the way we 

interpret and analyze visual data, becoming an essential tool 
across various fields, from healthcare diagnostics to 
autonomous systems. By transforming raw images into 
valuable information, image processing techniques allow 
machines to perceive, interpret, and act on visual inputs, 
pushing the boundaries of innovation. With a blend of 
mathematics, algorithms, and creativity, this domain 
continuously opens up new possibilities for automating com‐ 
plex tasks and unlocking insights that are often invisible to 
the human eye. In several research that connect with 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

106 

 

Glycemic Index, Image Processing techniques were used for 
different purposes. The earliest work from our repository 
appears in 2023; A Novel Machine Learning and Deep 
Learning Driven Prediction for Pre‐diabetic Patients [30]. 
They present a Machine Learning and Deep Learning‐ based 
prediction model that predicts the Glycemic Index of fruits, 
specifically bananas, using image recognition and 
classification. The model aims to determine if a fruit is safe for 
consumption by pre‐diabetic patients based on its glycemic 
properties. The system also provides food recommendations 
with high dietary fiber to help users maintain a balanced diet. 
From the point of image processing, techniques such as 
binarization and boundingRect from OpenCV are used for 
predicting the length of fruits. These images were then used 
to train a CNN for fruit recognition and classification. 
Similarly, aiming at diabetic patients [93], focuses on 
recognizing fruits and estimating the GL index. The primary 
goal is to help diabetic individuals make informed decisions 
about their daily diet by identifying fruits with high 
sweetness but low GL values. The Glycemic Index is indirectly 
used in estimating the GL of fruits. The GL considers both the 
GI value and the carbohydrate content of the food, providing 
a measure of how a particular food affects blood sugar levels. 
By identifying fruits and estimating their volume, the study 
calculates the GL index, helping diabetic patients determine 
whether a fruit is suitable for consumption based on its 
expected impact on blood glucose levels. The authors 
developed a custom dataset called DODP, containing 54,000 
images of fruits captured from various angles and 
environments where image preprocessing and data 
augmentation have been utilized. Automatic White Balance 
(AWB) and histogram equalization were used to improve the 
color and contrast of the images to improve the consistency 
of the input images. To generalize the images, data 
augmentation techniques such as adding Gaussian white 
noise, Pretzel noise, image flipping, rotation, and panning 
were applied, making a more diverse dataset for training. 
Going along the same direction of implementing data sets, 
[11], have created and used a Moroccan food dataset 
containing 72 dishes and 8,300 images. The primary goal is to 
use this system for nutritional analysis, for estimating the 
Glycemic Index and GL of Moroccan dishes, for dietary 
planning and management for chronic conditions such as 
diabetes. As the main AI technology, DenseNet deep learning 
models with an attention mechanism were used to improve 
the accuracy of food image classification. From the image 
processing perspective, mainly as in the previous, data 
augmentation has been done with the use of techniques like 
flipping, rotation, cropping, and noise removal in order to 
expand the dataset and improve model generalization.  

Another recent research work presents a system that 
recognizes food items from images uploaded by users and 
predicts their nutritional values, including GI, proteins, 
carbohydrates, and fats [25]. They have used a custom 
Inception‐V3 model for food image recognition and 
classification. As images are the main source, and users are 
responsible for uploading them, different image processing 
techniques such as noise reduction, histogram equalization, 
and data augmentation were utilized. Specifically, a 3x3 
median filter is applied to reduce “salt and pepper” noise, 
Contrast Limited Adaptive Histogram Equalization (CLAHE) 
is used to standardize lighting and color profiles, improving 
generalization across varied food images. Apart from them, 
data augmentation techniques such as shearing, zooming, 
rotation, and horizontal flipping are used to artificially 
expand the dataset for smoother training. 

In addition to such direct use of Image Processing 
techniques, we found one review article summarizing the 
current state of mobile‐based Food Image Recognition 
Systems (FIRS) designed for the dietary management of 
diabetics [95]. The focus is on evaluating technologies for 
classifying food, estimating food volume, and calculating the 
nutritional content of foods using smartphone cameras and 
computer vision techniques. The paper reviews approaches 
for using deep learning, machine learning, and image 
processing to automate food classification and support 
diabetic diet management. In the presented papers, the 
following image processing techniques were used 
appropriately. 
• Preprocessing: Enhancing food images by correcting 

lighting, scaling, cropping, and applying contrast 
adjustments. 

• Segmentation: Separating different food items within an 
image using methods such as manual segmentation, 
thresholding, color/texture‐based segmentation, CNN‐
based segmentation, and clustering‐based techniques. 

• Feature extraction: Extracting visual features from the 
images, such as color, texture, shape, and edges, using 
methods like Scale‐Invariant Feature Transform (SIFT), 
Histogram of Oriented Gradients (HOG), Gabor filters, and 
Local Binary Patterns (LBP). 

• Volume estimation: Using geometric modeling, pixel 
counting, and 3D reconstruction techniques from multi‐
view images to estimate the volume of food items. Depth 
map fusion techniques and shape‐fitting methods (e.g., 
cylinders, spheres) are also reviewed. 

Although image processing techniques were not the primary 
focus of the research discussed, they emerged as essential 
tools for facilitating the effective application of other AI 
techniques. Image processing techniques have contributed to 
enhancing the quality of AI-powered systems that predict GI 
values or dietary plans by transforming images into more 
insightful information sources. Image processing techniques 
such as noise reduction and histogram equalization were 
employed to enhance the images. 

4.5 Role of natural language processing (NLP) in 
research related to the Glycemic index 
Natural Language Processing, which reshapes how 

researchers handle vast amounts of unstructured text, 
enables deeper insights and more efficient data 
interpretation. The subject has transformed the way we 
understand and interact with text‐based data, becoming a 
crucial component in fields ranging from healthcare to 
customer service automation. By converting raw text into 
structured information, NLP techniques empower machines 
to comprehend, interpret, and generate human language, 
enhancing our ability to analyze large volumes of textual 
content. With the fusion of linguistics, machine learning, and 
algorithms, NLP is unlocking new opportunities to automate 
tasks such as translation, sentiment analysis, and information 
retrieval, offering deeper insights into language patterns that 
are often difficult for humans to detect. In many studies 
related to sentiment and behavior analysis, NLP techniques 
have been applied for various purposes, demonstrating their 
growing importance across disciplines. Research that is based 
on Glycemic index-related studies also benefited from this 
technology and has opened many research avenues as well. 
One of the first studies in 2019 related to NLP explores the 
estimation of the glycemic impact of cooking recipes using a 
data‐driven approach, combining online crowd‐sourcing and 
machine learning [72]. The goal was to classify recipes as 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

107 

 

healthy or unhealthy for diabetics based on their glycemic 
impact. They used Amazon Mechanical Turk (AMT) workers 
to crowdsource glycemic impact labels for recipes and used 
machine learning models to predict whether a recipe was 
unhealthy for diabetics (UD) or healthy (HD). The study used 
a data set of 55,102 All-recipes recipes, narrowing it down to 
990 recipes (Recipe990) for a detailed analysis based on the 
range of glycemic impact and the difficulty of classification. 
AMT was used to gather human judgments on the glycemic 
impact of these recipes. Due to the impracticality of using the 
Glycemic Index for large datasets, the researchers relied on 
the sugar‐to‐fiber ratio (S/F) as a proxy for the glycemic 
impact. To process textual data, NLP techniques such as Bag‐
of‐Words, word embeddings (word2vec, GloVe, fastText), 
sentence embeddings (skip‐thought vectors), and paragraph 
embeddings (doc2vec) were used. The best performance 
among models using only textual features came from the NB‐
BoW + Logistic Regression model with an F1 score of 0.817. 
When nutritional features were added, the GloVe + LightGBM 
model achieved the highest overall F1 score of 0.854, 
highlighting that the combination of textual and nutritional 
data improves the precision of estimating the impact of 
glycemics for diabetics. In addition to the primary focus above 
on dietary management technologies, one review article 
systematically evaluates mobile‐based Food Image 
Recognition Systems aimed at dietary management for 
diabetics [95]. Relevant articles published over the last two 
decades are evaluated. The paper discusses the importance of 
integrating the Glycemic Index and GL into food classification 
systems, which is crucial to predicting the impact of food on 
blood glucose levels. It emphasizes that diabetic patients can 
benefit from technologies that predict GI and GL. Although the 
review discusses the potential use of GI in future applications, 
it primarily reviews current methods that focus on food 
recognition and volume estimation, without directly using GI 
data in the reviewed systems. The paper assesses various 
methodologies for classifying food, estimating portion sizes, 
and calculating nutritional content using smartphone 
cameras and advanced NLP techniques. 

These papers mainly follow some of the NLP techniques 
and concepts with the goal of improving food classification 
and nutritional estimation. 
• Textual analysis: Evaluating the nutritional information 

associated with food items through text data from various 
sources, such as recipes and nutritional databases. This 
involves extracting relevant information about 
macronutrients, ingredients, and portion sizes from 
written content. 

• Ingredient recognition: Using NLP techniques to parse and 
recognize food ingredients from text descriptions, which 
could complement image classification by providing 
additional context regarding food items that may not be 
visually distinguishable. 

• Recipe parsing: Developing methods to extract nutritional 
data and estimates from ingredient lists found in recipes. 
This includes recognizing quantities and types of 
ingredients to calculate potential carbohydrate content, 
which is crucial for diabetic diet management. 

• Machine learning for text classification: Employing text 
classification algorithms (e.g., logistic regression, SVM) to 
categorize recipes or food items based on their healthiness 
scores or glycemic impact, derived from textual data 

• Textual data extraction: NLP techniques were utilized to 
extract and process textual information from food labels, 

menus, or recipes associated with food images, enabling a 
more comprehensive analysis of nutritional content. 

NLP techniques have been used to analyze text-related data 
such as recipes, nutritional descriptions, and ingredient lists 
to extract features that require further processing to reveal 
insights such as carbohydrate content, sugar-to-fiber ratios, 
and overall dietary healthiness. These results can be 
combined with visual features for more accurate glycemic 
impact prediction. Beyond classification, NLP also supports 
ingredient recognition, recipe parsing, and textual data 
extraction from diverse sources, enabling automated and 
scalable dietary assessment. 

 

        2024 ‐ 2015   [11],[17],[20][21],[22],[27],[29], 

   [30],[31],[32],[33],[38],[39],[63], 

   [65],[66],[67],[69],[71],,[72],[72], 

   [73],[73],[74],[74],[75],[77], 

   [78],[80],[83],[84],[85],[87],[88], 

   [88],[91],[92],[93],[93],[94],[95],  

   [95],[95],[97],[98],[99],[100], 

  [101],[102],[111] 

 

2014 ‐ 2005 [44],[103],[106],[107]  

 

2004 ‐ 1995 [108]  

 

1994 ‐ 1985    [109],[110] 

  

1984 ‐ 1980 
 

Figure 7: Timeline of research with categories; Background colors 
represent different categories of researches: 
Machine Learning , Deep Learning , NLP ,Reinforcement Learning , 
Image Processing , Statistical Learning , Explainable AI , Others 

4.6 Role of explainable AI (XAI) in research related to the 
Glycemic index 
Only a few studies in our repository have directly applied 

Explainable AI (XAI) techniques in their research. The earliest 
example identified was published in 2023 [112]. This study 
employed Shapley Additive exPlanations (SHAP), a widely 
used XAI method, to evaluate the influence of various meal‐
related factors on predicting postprandial blood glucose 
levels at different time intervals. SHAP assigns importance 
values to the features, effectively highlighting their 
contributions to the predictions of the model. By utilizing 
SHAP, the researchers provided valuable insights into the 
effects of specific nutritional components, such as 
carbohydrate intake, protein, lipids, and Glycemic Index, on 
blood glucose levels in individuals with type 1 diabetes. This 
approach not only validated clinical hypotheses but also 
enhanced the interpretability of predictive models, fostering 
more transparent and informed decision‐making in diabetes 
management. Another research has used SHAP to enhance the 
interpretability of its predictive models [53]. After developing 
a Diabetic Retinopathy (DR) risk prediction model using the 
CatBoost algorithm, the researchers applied SHAP to 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

108 

 

interpret the model’s outputs. SHAP values helped in 
understanding the contribution of each feature to the model’s 
predictions, thereby elucidating the relationship between 
various risk factors and the likelihood of developing DR. SHAP 
was employed to rank features based on their impact on the 
prediction model. This ranking identified which factors were 
most influential in predicting DR risk, providing insights into 
the relative importance of different clinical indicators. Also, 
the study used SHAP to analyze the correlations between 
renal function indices and other measures. This analysis 
offered a deeper understanding of how different health 
indicators interact and contribute to the risk of developing 
DR. The integration of SHAP into the predictive modeling 
process revealed several key risk factors for diabetic 
retinopathy: 
• Positive correlations: Factors such as albumin‐to‐

creatinine ratio (ACR), glycated hemoglobin, 24‐hour 
urinary protein, presence of nephropathy, and serum 
creatinine (SCR) were found to be positively correlated 
with DR. This means that higher values of these indicators 
are associated with an increased risk of developing DR. 

• Negative correlations: Conversely, factors like C‐peptide 
(CP), hemoglobin (HB), albumin (ALB), direct bilirubin 
(DBILI), and C‐reactive protein (CRP) were negatively 
correlated with DR, indicating that higher levels of these 
factors might be associated with a reduced risk. 

• Non‐significant factors: The study found that 
characteristics such as height, weight, and erythrocyte 
sedimentation rate (ESR) did not have a significant 
relationship with the development of DR. 

By employing SHAP, the researchers were able to provide a 
transparent and interpretable analysis of the predictive 
model, making the findings more understandable and 
actionable for clinical applications. This approach improves 
the reliability of the prediction results and helps identify 
critical factors for early prevention and clinical diagnosis of 
diabetic retinopathy. These two significant contributions, 
incorporating the latest machine learning advancements such 
as Explainable AI methods, highlight that the path is open for 
exploring Explainable AI in future research on the Glycemic 
Index. Using XAI techniques in glycemic and diabetes-related 
research has bridged the gap between model accuracy and 
interpretability. Past studies have shown that XAI techniques 
can be used to unveil nutritional and clinical factors 
influencing blood glucose dynamics and diabetic 
complications. XAI enables researchers to quantify each 
feature’s contribution to predictions, thereby improving trust 
in AI-assisted decision-making. Although only a few studies 
have incorporated XAI techniques, these studies have laid a 
strong foundation for future research to use XAI methods 
more extensively to analyze predictions. 

4.7 Role of statistical techniques in research related to 
the Glycemic index 
 Statistical Techniques are essential tools for analyzing 

data and drawing conclusions in various fields, including 
social sciences, health, and business. These methods can be 
categorized into descriptive statistics, which summarize and 
visualize data through measures such as mean, median, and 
mode, and inferential statistics, which enable researchers to 
make predictions or generalizations about a population based 
on sample data. Key components of statistical techniques 
include hypothesis testing, where null and alternative 
hypotheses are formulated to evaluate the significance of 
results using p‐values and confidence intervals; correlation 
and regression analysis, which examine relationships 

between variables to determine how one may predict or 
influence another; and t‐test and ANOVA, which compare 
means across different groups to assess whether observed 
differences are statistically significant. Overall, statistical 
techniques provide a robust framework for making informed 
decisions based on empirical evidence, ensuring that 
conclusions drawn from data are reliable and valid. 

Most probably, the earliest work of such appears in 
1990: “Glycemic Index of Foods in Individual Subject” [109]. 
The study involved 12 diabetic subjects consuming mixed 
meals (bread, rice, spaghetti) to determine common Glycemic 
Index values. The researchers calculated GI based on the area 
under the glycemic response curve, using white bread as a 
reference. They used analysis of variation with repeated 
measures (ANOVARM) to assess differences in glycemic 
responses between meals and subjects, along with the 
Tukey’s Q method for adjustment of multiple comparisons 
and Chi‐square analysis to compare observed rankings with 
expected outcomes. The findings indicated that while 
individual responses varied, mean GI values for each food 
type were consistent across subjects, validating the predictive 
capability of GI in dietary studies among diabetic patients. 
Building on this foundation, in 1993, “Prediction of Glycemic 
Index for starchy foods” [113] analyzed 18 starchy foods to 
identify predictive factors for GI based on food components 
such as protein, fat, and total dietary fiber (TDF). GI was 
calculated using the area under the glucose response curve 
for each food relative to white bread. Statistical techniques 
included Regression Analysis to explore correlations between 
GI and food components, t‐tests for comparing means 
between legumes and non‐legumes, and calculating 
Correlation Coefficients to quantify relationships between GI 
and food components. The study highlighted that while 
certain food components correlate with GI, preparation 
methods and starch characteristics significantly influence 
glycemic responses. 

A study by M. Mayo et al. [73] investigated the thresholds 
of fasting blood glucose and glycosylated hemoglobin 
associated with microalbuminuria. They enrolled 975 
subjects, including 873 diabetic patients and 102 non‐
diabetic controls, to analyze the impact of glycemic control on 
microalbumin levels. The study does not specifically use the 
Glycemic Index but focuses on glycemic control measured by 
FBS and HbA1c levels. Explore how poor glycemic control 
affects microalbuminuria, which is relevant to understanding 
the impacts of diet on health. They used the Student t‐test for 
comparing means between two groups and the Analysis of 
Variance to assess differences among multiple groups. 
Multiple Linear Regression is used to identify the relationship 
between variables and urinary microalbumin levels. Chi‐
squared analysis was employed for comparing prevalence 
rates. These techniques were used to evaluate differences in 
FBS, HbA1c, and homocysteine levels among different groups 
and to develop predictive models for microalbuminuria. 

Further expanding the understanding of GI, the [108] 
study on composite breakfast meals involved 28 healthy 
young men testing 13 different meals. Researchers used 
Regression Analysis and Multivariate Analysis to develop 
prediction equations for GI based on meal components. The 
findings indicated that energy density and fat/protein ratios 
were more reliable predictors of GI than carbohydrate 
content alone. Together, these studies illustrate how 
statistical techniques are crucial for accurately determining 
and predicting glycemic responses, ultimately aiding in better 
dietary choices for individuals managing blood sugar levels. 
The study by Mohan et al. [74] explores the association 

https://docs.google.com/document/d/14wrzRJbMmXrNMvtgEGZciuAlQ2bP6c8u/edit#heading=h.3q5srcqew924
https://docs.google.com/document/d/14wrzRJbMmXrNMvtgEGZciuAlQ2bP6c8u/edit#heading=h.x7rvsuq2zx
https://docs.google.com/document/d/14wrzRJbMmXrNMvtgEGZciuAlQ2bP6c8u/edit#heading=h.rsfuub18cwjz
https://docs.google.com/document/d/14wrzRJbMmXrNMvtgEGZciuAlQ2bP6c8u/edit#heading=h.3q5srcqew924


NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

109 

 

between poor glycemic control, hyperhomocysteinemia, and 
microalbuminuria, utilizing both traditional statistical 
methods and machine learning algorithms. Although not 
directly focused on the Glycemic Index, this research 
highlights the importance of statistical techniques in 
understanding health outcomes related to glycemic control. 
The researchers focused on glycemic control as measured by 
fasting blood sugar and glycosylated hemoglobin levels. The 
main technologies used included biochemical analysis 
equipment and computational software for machine learning. 
The dataset consisted of 975 subjects, including diabetic 
patients and non‐diabetic controls, with features such as age, 
gender, FBS, HbA1c, and diabetic status. Statistical techniques 
employed included Student t‐tests, ANOVA, multiple linear 
regression, and Chi‐squared analysis to assess differences 
and model relationships between variables. 

Another study introduces a hybrid model designed to 
predict GI and GL based on the macronutrient composition of 
foods [33]. This model combines deterministic calculations 
for glycemic carbohydrates with empirical coefficients for 
non‐glycemic nutrients like proteins, fats, and fibers. By 
quantifying both the impact of glycemic carbohydrates and 
the GI‐lowering effects of non‐glycemic components, the 
model aims to facilitate the development of packaged foods 
and beverages with lower glucose responses. The model was 
validated using a dataset of 42 breakfast cereals and 60 in 
vivo trials, employing statistical techniques like Ordinary 
Least Squares (OLS) regression and Bland‐Altman plots to 
achieve high correlation coefficients (r = 0.90 for GI and r = 
0.96 for GL). The transparency of the model, with explicit 
coefficients for each nutrient, makes it interpretable and 
useful to guide the development of the product. 

The study “Gluten‐free cookies with low Glycemic Index 
and glycemic load: Optimization of the process variables via 
response surface methodology and artificial neural network” 
[80] focused on optimizing the production of gluten‐free 
cookies with low Glycemic Index and glycemic load. It 
evaluates the impact of baking temperature and time on 
resistant starch (RS), GI, and GL using cardaba banana flour 
modified with citric acid to enhance RS content. The study 
employs Response Surface Methodology (RSM) and Artificial 
Neural Networks to model and optimize these parameters. 
RSM is used to understand the relationships between baking 
conditions and outputs, while ANN provides more accurate 
predictions. The GI is estimated through in vitro starch 
digestibility tests using a non‐linear model. Experimental 
data from 13 baking trials, designed using a central composite 
design, were analyzed using techniques like ANOVA and 
regression analysis to assess model quality and optimize 
baking conditions. Overall, the study aims to develop gluten‐
free cookies with improved nutritional profiles by optimizing 
production parameters. 

The study [83] uses historical glucose data from 
continuous glucose monitoring devices to predict future GI 
levels, with Autoregressive Integrated Moving Average 
adapting in real‐time and RNN predicting trends over 30‐60 
minutes. The models are hosted on Google Cloud, utilizing 
technologies like Google Cloud PubSub, Functions, and 
BigQuery for real‐time training and data management. The 
D1NAMO dataset, which includes glucose readings from 
diabetic and non‐diabetic patients, is used for model training 
and validation. Statistical techniques such as ADF tests, 
ACF/PACF, and Akaike Informa‐ tion Criterion are employed 
for model optimization and validation. The study aims to 
monitor glucose levels in bus drivers, providing alerts for 

dangerous trends, and exploring the inherent interpretability 
of ARIMA models to understand forecast dependencies. 

Reference [38] focused on determining the Glycemic 
Index of a complete nutrition drink formulated with 
retrograded starch and identifying factors influencing the 
glycemic response. This was achieved through a randomized 
crossover controlled trial involving 18 healthy participants 
who consumed the nutrition drink, glucose solution, and 
white bread as test foods. Normality tests, such as the 
Shapiro‐Wilk test, are employed to evaluate the distribution 
of data. For comparative analysis, a one‐way ANOVA is used 
for normally distributed data, while the Kruskal‐Wallis and 
Friedman tests are applied for non‐parametric comparisons. 
In scenarios involving repeated measures, repeated measures 
ANOVA with Tukey’s test for post‐hoc analysis is performed. 
For correlation analysis, the Spearman rank correlation is 
utilized to identify relationships between baseline 
characteristics and glycemic response. To control Type I 
errors in multiple comparisons, the Bonferroni correction is 
applied. Statistical techniques were primarily used to 
compare postprandial glucose and insulin levels across test 
foods, analyze baseline characteristics among groups, and 
identify correlations between predictors (e.g., baseline 
insulin) and glycemic response (Table 2). 

“Predicting Changes in Glycemic Control Among Adults 
with Prediabetes from Activity Patterns Collected by 
Wearable Devices” [88]. The study explores the use of 
wearable devices to predict changes in glycemic control 
among adults with prediabetes, focusing on comparing wrist-
worn and waist‐worn devices. Participants were monitored 
over six months using Fitbit devices that tracked physical 
activity, sleep, and heart rate. The study did not involve 
calculating the GI but instead focused on predicting changes 
in hemoglobin A1c using wearable data and machine learning 
models. Traditional statistical regression models and 
machine learning techniques like random forest and 
ensemble methods were employed to analyze baseline 
demographic, clinical, and wearable data. Features included 
demographics, clinical data, physical activity, heart rate, and 
sleep patterns, which were reduced to 16 principal 
components using PCA. The study used techniques such as 
multiple imputations for missing data and hyper‐parameter 
tuning with cross‐validation. While transfer learning and 
explainable AI were not explicitly used, ensemble methods 
provided insights into predictive factors. The goal was to 
enhance predictive models for glycemic control changes, 
leveraging wearable data to potentially inform interventions 
for preventing diabetes progression. 

Enhancing insights into GI, the study [17] aimed to 
enhance the Venezuelan Food Composition Table by 
integrating Glycemic Index values to aid in dietary 
assessments and research. It employed a systematic six‐step 
methodology to assign GI values to 624 food items across 14 
categories. This approach included direct assignment from 
international GI tables, mapping to similar foods, recipe‐ 
based calculations, and using subgroup median values for 
unassignable items. Key features influencing GI assignments 
were available carbohydrates, nutritional profiles (including 
fat, protein, and fiber), food preparation methods, and other 
nutrient compositions. The study utilized IBM SPSS for data 
analysis and adhered to the ISO 26642:2010 standard for GI 
determination. Statistical analyses were utilized to stratify 
the results by food group and to perform calculations such as 
mean, standard deviation, and percentile distribution of 
Glycemic Index.  

 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

110 

 

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
These analyses also facilitated the classification of foods 

into low, moderate, or high GI categories. Descriptive 
statistics were applied to summarize the GI distributions, 
while proportional analysis was employed to determine the 
percentage of foods assigned GI values at each step of the 
methodology. Another study investigates the relationship 
between blood glucose data and various physiological and 
nutritional factors using wearable devices and publicly 
available datasets. It utilizes devices like the Dexcom G6 CGM 
for blood glucose monitoring and the Empatica E4 Wristband 
for capturing physiological signals such as heart rate and skin 
temperature.  

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
The analysis employs software tools like SAS 9.4, JMP Pro 

16.1.0, and Microsoft Excel, and statistical techniques 
including correlation analysis, multiple regression, and one‐
way ANOVA. The study draws data from the PhysioNet BIG 
IDEAs Lab Dataset, which includes 16 participants with 
prediabetic HbA1c levels, focusing on glucose, physiological 
indices (e.g., heart rate, skin temperature), and nutritional 
indices (e.g., carbohydrates, dietary fiber). Notably, the study 
does not directly use or predict the Glycemic Index; instead, it 
focuses on real‐time blood glucose data collected via CGM 
devices to explore correlations between physiological and 
nutritional indices and blood glucose levels, as well as assess 

Table 2. Summary of statistical techniques in the glycemic index research 

Paper   Reference Statistical Tools Used Purpose of Statistical Test 

[109] ANOVA with Repeated Measures 
(ANOVARM), Tukey’s Q method, Chi-square analysis 

Assess differences in glycemic responses between meals 
and subjects; adjust for multiple comparisons; compare 
observed rankings with expected outcomes. 

[113] Regression Analysis, t‐tests, Correlation Coefficients Explore correlations between GI and food components; compare 
means between legumes and non‐legumes; quantify 
relationships between GI and food components. 

[73]  Student t‐test, Analysis of Variance, 
Multiple Linear Regression, Chi-squared analysis 

Compare means between groups; assess differences 
among multiple groups; identify relationships between variables 
and urinary microalbumin levels; compare 
prevalence rates. 

[108] Regression  Analysis, Multivariate Analysis Develop prediction equations for GI based on meal components; 
identify reliable predictors of GI. 

[74]  Student t‐tests, ANOVA, Multiple Linear Regression, 
Chi‐squared analysis 

Assess differences and model relationships between 
variables; explore associations between glycemic control, 
hyperhomocysteinemia, and microalbuminuria. 

[76]  Ordinary Least Squares Regression, 
Bland‐Altman plots 

Develop a hybrid model to predict GI and GL based on 
macronutrient composition; validate model accuracy. 

[80]  Response Surface Methodology, Artificial Neural 
Networks, ANOVA, Regression Analysis 

Optimize production of gluten‐free cookies with low GI 
and GL; model and optimize baking conditions; assess 
model quality. 

[83]  Autoregressive Integrated Moving 
Average, Recurrent Neural Networks (RNN), ADF 
tests, ACF/PACF, 
Akaike Information Criterion 

Predict future GI levels using historical glucose data; 
adapt models in real‐time; optimize and validate models. 

[38]  Shapiro‐Wilk test, One-way ANOVA, 
Kruskal‐Wallis test, Friedman test, Repeated 
Measures ANOVA, Tukey’s test, Spearman Rank 
Correlation, 
Bonferroni Correction 

Evaluate data distribution; compare postprandial glu‐ 
cose and insulin levels across test foods; analyze baseline 
characteristics; identify correlations between predictors and 
glycemic response; control Type I errors in multiple 
comparisons. 

[88]  Principal Component Analysis, 
Random Forest, Ensemble Methods, Multiple 
Imputation, Hyper‐parameter Tuning, Cross-
validation 

Predict changes in glycemic control using wearable data; 
enhance predictive models; manage missing data; optimize 
model performance. 

[17]  Descriptive Statistics, Proportional 
Analysis 

Stratify results by food group; calculate mean, standard 
deviation, and percentile distribution of GI; classify foods into GI 
categories; determine the percentage of foods 
assigned GI values at each methodological step. 

[21]  Correlation Analysis, Multiple Regression, One‐way 
ANOVA 

Explore relationships between blood glucose data and 
physiological/nutritional factors; assess postprandial glucose 
dynamics; analyze data from wearable devices. 

[22]  Additive Main Effect and Multiplicative Interaction 
(AMMI) Analysis, GGE Biplot Analysis, Linear Mixed 
Models (LMM) 

Evaluate grain yield, quality traits, and genotype‐environment 
interactions; predict GI in rice varieties; develop low‐GI rice 
suitable for specific ecosystems. 
  

 

 

 

 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

111 

 

postprandial glucose dynamics [21]. Extending expertise on 
GI, Reference [22] aimed to develop low‐glycemic‐index rice 
varieties suitable for irrigated ecosystems in Bangladesh. It 
evaluated the grain yield, quality traits, and genotype‐
environment interactions (GEI) of three rice genotypes across 
27 environments over three years. The GI values were 
determined using the Howlader and Biswas technique, which 
involves measuring postprandial blood glucose levels after 
consuming test and reference foods. Statistical analyses 
included Additive Main Effect and Multiplicative Interaction 
analysis, GGE biplot analysis, and Linear Mixed Models (LMM) 
using R software. Features such as milled rice outturn, head 
rice yield, amylose content, and protein content were used to 
predict GI. This research is crucial for developing rice 
varieties that can benefit health by reducing the impact of 
carbohydrates on blood glucose levels. 

Overall, statistical techniques play a pivotal role in 
understanding and predicting glycemic responses, which are 
crucial for managing blood sugar levels. These methods 
encompass descriptive and inferential statistics, hypothesis 
testing, correlation and regression analysis, and tests like t‐
tests and ANOVA. Studies have utilized these techniques to 
determine Glycemic Index values for various foods, identify 
predictive factors for GI, and develop models to predict GI 
based on food components. For instance, regression analysis 
has been used to explore correlations between GI and food 
components, while response surface methodology and 
artificial neural networks have been employed to optimize 
production parameters for low‐GI foods. Additionally, 
machine learning and wearable devices are being explored to 
predict changes in glycemic control, further enhancing the 
application of statistical techniques in this field. These 
methods provide a robust framework for making informed 
dietary choices and developing healthier food products. 

4.8 Role of other computer-related techniques in 
research related to the Glycemic index 
In the main section, we explored the role of AI techniques 

such as Machine Learning, Deep Learning (DL), Image 
Processing, and Natural Language Processing in research 
related to the Glycemic Index. While these advanced 
methodologies have transformed the field, there are other 
computer‐related techniques that also contribute 
significantly to Glycemic Index research. This section focuses 
on these additional approaches, highlighting their unique 
applications and the value they bring to advancing our 
understanding of this critical area. The Internet of Things 
(IoT) has played a pivotal role in predicting the Glycemic 
Index. IoT refers to a network of interconnected physical 
devices ranging from appliances to vehicles embedded with 
sensors, software, and connectivity. This technology 
facilitates seamless communication and data exchange 
between devices, paving the way for more efficient and 
automated systems. In recent years, challenges such as data 
mining, machine learning integration, and IoT applications 
have gained prominence in the healthcare sector. A notable 
study published in 2019, titled “Internet of Things Based on 
Electronic and Mobile Health Systems for Blood Glucose 
Continuous Monitoring and Management” [69], showcased 
the integration of the Libre flash glucose monitoring sensor 
with mobile applications, creating a connected and 
comprehensive environment for glucose monitoring. They 
used cloud technologies to collect blood glucose data 
continuously, provide real‐time alerts, and perform graphical 
analysis while monitoring and analyzing patient data 
remotely via a secure cloud‐based platform. However, GI was 

not a direct concern; the system focused on real‐time 
monitoring of blood glucose. These data were used to identify 
patterns related to GI. In general, the study emphasizes the 
usage of IoT in the health care system to manage diabetes and 
low‐cost alternatives to traditional methodologies for 
continuous glucose monitoring systems. The utilization of the 
long‐term effect of the Internet of Things on glycemic control 
is controversial, and Type 2 diabetes is a common problem 
today. Another study focused on evaluating the long‐term 
effects of an IoT‐based approach on glycemic control in 
people with Type 2 Diabetes (T2D). The personal health 
records (weights, blood pressure, physical activities) were 
measured using IoT-enabled devices, and feedback messages 
were sent to encourage behavioral changes in diet and 
exercise. Data was shared with healthcare providers via cloud 
systems. GI was not directly addressed; instead, it focused on 
glycemic control through HbA1c levels and lifestyle 
modifications facilitated by IoT technologies [100]. The 
advancements in wearable glucose monitoring technologies, 
as reviewed by Mansour et al. [101]. This offers significant 
implications for Glycemic Index research. The paper reviewed 
advancements in wearable devices for CGM, including 
invasive, minimally invasive, and non‐invasive methods. The 
paper highlights the integration of biosensing technologies 
with wireless communication, energy harvesting, and AI‐
based predictive analytics for diabetes management. The 
Glycemic Index is not explicitly discussed or used for 
predicting or calculating glucose levels. Instead, the focus was 
on measuring glucose directly from biofluids (e.g., blood, 
sweat, interstitial fluid) using various biosensor technologies. 
By integrating AI‐driven models like RNNs, researchers can 
better account for factors such as physical activity, stress, and 
insulin sensitivity, enhancing the accuracy of GI predictions. 
The insights from this review pave the way for a more 
personalized and scalable approach to the management of 
diabetes and diet.  

A healthy and balanced diet is essential for quality of life. 
Carbohydrates play a crucial role in maintaining a healthy and 
balanced diet, since they serve as the primary source of 
energy in the body. Going along with Mathematical 
approaches, the study, “A robust optimization approach to 
diet problem with overall GL as objective function” [106], 
addresses the problem of minimizing the overall GL in daily 
diets while meeting nutritional and serving size 
requirements. The authors proposed a mixed‐integer 
programming model that incorporates uncertainties in GL 
values, allowing for flexible and adaptive diet planning. This 
study focuses on creating a mathematical framework to 
optimize daily food selection while minimizing the total GL. 
This optimization ensures that daily nutritional needs are 
met, minimizes the impact of foods with a high GL on blood 
glucose levels, and allows for uncertainty in GL values so that 
meals can be maintained under different circumstances. 
Another study investigates the effects of dietary Glycemic 
Index on β‐cell function in adults with prediabetes through a 
randomized controlled feeding trial. A total of 35 adults with 
prediabetes underwent a 2‐week control diet (GI = 55–58), 
followed by randomization into a 4‐week low Glycemic Index 
(LGI; GI < 35) or high Glycemic Index (HGI; GI > 70) diet. Meals 
were carefully designed to meet GI specifications while 
maintaining consistent macronutrient distribution (55% 
carbohydrate, 30% fat, 15% protein) and ensuring weight 
stability. Meal tolerance tests (MTTs) were conducted at 
baseline and post‐intervention to evaluate glucose, insulin, 
and C‐peptide responses. Mathematical models and statistical 
tools (e.g., SPSS, MATLAB) were used to estimate β‐cell 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

112 

 

glucose sensitivity, insulin secretion rates, and insulin 
sensitivity indices (OGIS, Matsuda index). The LGI diet 
significantly reduced postprandial glucose concentrations (p 
< 0.001) and increased total insulin secretion adjusted for 
glucose levels and insulin sensitivity (p = 0.002). Conversely, 
the HGI diet showed trends towards higher glucose levels (p 
=0.14) and reduced insulin secretion. Despite these 
differences, neither diet significantly affected traditional 
measures of insulin sensitivity. These findings highlight the 
potential of LGI diets in improving β‐cell function and glucose 
regulation in individuals with prediabetes [32]. 

Continuous glucose monitoring is an important aspect 
for diabetic patients. There are many studies that focus on 
that in different ways. A study aimed to automate the 
calculations of the GI using a continuous glucose monitoring 
system (CGMS). GI was calculated using the incremental area 
under the blood glucose curve (IAUC). They calculated the 
IAUC using CGMS. A custom-built Microsoft Excel-based 
software for the automation process was introduced. This 
software, called DegifXL, processed CGMS data, calculated 
IAUC, and computed GI values using pre‐defined and custom 
input parameters. The process reduces the manual 
processing time and improves the standardization of GI 
computations [103]. 

The study “Beyond nutrient‐based food indices: a data 
mining approach to search for a quantitative holistic index 
reflecting the degree of food processing and including 
physicochemical properties” [65] explores the relationship 
between food processing, nutritional quality, and health 
potential, with a focus on the Glycemic Index as a key 
indicator of glucose bioavailability. Using data mining 
techniques such as decision trees, Bayesian networks, and 
Principal Component Analysis, the study analyzed 117 foods 
categorized by processing levels (minimally processed, 
processed, and ultra‐processed) to develop a holistic 
Technological Index (TI). This index integrates functional 
properties like nutrient density, glycemic glucose 
equivalents, and physicochemical characteristics such as 
texture and water activity. Results showed that minimally 
processed foods generally exhibit lower GI, higher nutrient 
density, and better satiety profiles compared to ultra‐
processed foods, emphasizing the adverse nutritional impact 
of food processing. The study highlighted how GI and other 
food properties can inform health-focused dietary guidelines 
and aid in promoting healthier food choices.  

Complementing this, the paper “AI4FoodDB: A database 
for personalized e‐Health nutrition and lifestyle through 
wearable devices and artificial intelligence” [92], established 
a comprehensive database integrating IoT‐enabled wearable 
devices, food diaries, and biological samples to explore 
relationships between diet, physical activity, and glycemic 
responses. While GI is not directly measured, continuous 
glucose monitoring data and dietary logs provide insights into 
postprandial glucose variability. Technologies such as FitBit 
Sense and FreeStyle Libre 2 sensors were used for real‐time 
data collection, with AI and machine learning employed for 
analysis across 10 domains, including biomarkers, nutrition, 
and gut microbiome. Similarly, another study developed a 
comprehensive Glycemic Index and GL database for the 
United States (U.S.) using NHANES data (1999–2018) to 
analyze dietary carbohydrate quality and its health 
implications [98]. AI models, specifically OpenAI’s pretrained 
embedding tools, were employed to assign GI values to over 
7,976 unique food codes, achieving 75% initial accuracy. But 
after manual review and adjustments based on expert 
knowledge, only 31.3% of the AI’s predictions were kept. Key 

databases, including the International Tables of GI [37] and 
the Diogenes Study [114], provided reference values. The 
dietary GL was calculated by combining carbohydrate content 
and GI values, while statistical and mathematical methods, 
such as weighted averages and residual adjustments, ensured 
robust data analysis. Trends in GI and GL were examined 
across demographics, highlighting disparities in diet quality 
by sex, race, education, and income. While transfer learning 
enabled efficient GI assignment, manual adjustments ensured 
accuracy and interpretability, making this database a critical 
resource for precision nutrition and public health research. 
Together, these studies underscore the importance of 
integrating glycemic metrics, food processing indices, and 
advanced technologies like AI and IoT to promote precision-
based nutrition and health monitoring. 

5. Datasets used in the Glycemic index-related research 

In the constantly evolving area of Glycemic index-related 
studies, the dataset a researcher chooses can significantly 
influence the study’s results and overall impact. Some 
researchers choose established benchmark datasets to 
maintain consistency and enable comparability across 
studies, while others create custom datasets to better align 
with their specific research goals. This chapter examines both 
approaches, emphasizing the importance and application of 
these diverse data sources in Glycemic Index research. 
Building on the AI‐driven techniques introduced in Chapter 4, 
here we discuss how careful data selection and preparation 
are essential for pushing the field forward. The graph below 
highlights the type of dataset utilized in each reviewed paper, 
categorizing them as either benchmark datasets (pre‐
existing) or self‐developed datasets (created specifically for 
the study). 

 
Custom  [11],[13],[14],[15],[18],[19] [20], 
Dataset  [22],[23],[24],[25],[26],[27],[28], 
(Self developed) [29],[31],[32],[34],[35],[36], [39], 

[40],[41][60],[61],[64],[65],[66], 
[67],[69],[71],[74],[75],[76],[77], 
[78],[79],[82],[83],[86],[87],[88], 
[89],[90],[92],[93],[94],[95],[100],    

  [101],[[102],[103],[104],[105], 
[106],[107],[108],[109],[113] 

 
Existing  dataset [17],[21],[30],[33],[62],[63],  
 (Benchmark)              [72],[73],[83],[91],[98],[99] 
 
Both  (Existing [37],[42],[43],[44],[97] 
& Custom) 
 

Figure 8: Dataset used by the research (Benchmark/ Custom dataset 
(Self-developed) 

5.1 Utilization of custom datasets in technological 
approaches for Glycemic index assessment 

As illustrated in the graph, the majority of studies utilize self‐
developed datasets, as these are tailored to the specific 
purpose of the research, resulting in improved model 
outcomes. As the earliest record in our repository [13] uses a 
custom-built dataset that consists of 62 commonly consumed 
foods and sugars. The foods were tested individually on 
groups of 5–10 healthy volunteers, totaling 34 individuals (21 
male, 13 female). The dataset includes the Glycemic Index of 
various food items, calculated by measuring blood glucose 
levels over two hours after consuming the foods. The 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

113 

 

glycemic response was expressed as a percentage of the blood 
glucose area under the curve compared to an equivalent 
amount of glucose. The study aimed to classify foods based on 
their physiological effects on blood glucose levels to better 
guide dietary choices, particularly for diabetics. Our second-
earliest research [109] involved 12 diabetic subjects who 
were given three types of mixed meals: bread, rice, and 
spaghetti, with Glycemic index values of 100, 79, and 61, 
respectively. These meals were tested in a randomized block 
design, with each subject consuming each meal four times. 
The dataset includes glycemic response measurements over 
time, expressed as incremental areas under the glycemic 
response curves. The study investigated how well the GI 
values predicted glycemic response rankings for individuals. 
By normalizing responses to a standard reference (bread), 
the researchers showed that GI values could effectively rank 
glycemic responses across different subjects, despite 
individual variability.  

A particular research [14], done for starch‐rich foods 
typically consumed in the Mediterranean region with type 2 
diabetic patients, measuring the plasma glucose response 
after consumption of different food portions containing 50 g 
of available carbohydrates (spaghetti, white bread, potatoes, 
pizza, potato dumplings, and hard toasted bread). The 
researchers measured the glycemic response of these foods 
in terms of blood glucose levels over a specific postprandial 
period, providing insights into the dietary effects of these 
foods on glycemic control. For example, they found that 
spaghetti and potato dumplings had lower glycemic 
responses compared to bread and potatoes, attributed to 
differences in food structure and preparation techniques. 
Reference [15] prepared and analyzed Amaranthus cruentus 
seeds using various processing methods such as cooking, 
popping, roasting, flaking, and extrusion. The seeds were 
sourced from a local producer in Brasília, Brazil. Using the 
Hydrolysis Index (HI) derived from in vitro tests, they 
determined the Predicted Glycemic Index (pGI) for each 
processed seed sample. But in here, a glycemic response 
comparison has been conducted with white bread as a 
benchmark (reference sample). Reference [16] involved 28 
laboratories testing the Glycemic Index of foods such as 
cheese puffs and fruit leather. Each laboratory followed a 
standardized protocol where 10 healthy participants 
consumed test foods and reference foods (glucose or white 
bread) on separate occasions. Blood glucose responses were 
measured at regular intervals, and the Glycemic Index was 
calculated based on the incremental area under the curve of 
glucose response.  

Reference [40] aimed to determine the Glycemic Index 
values of traditional foods and mixed meals from Northern Sri 
Lanka. They conducted experiments to measure the GI values 
of various traditional foods and mixed meals consumed in 
Northern Sri Lanka. This involved selecting specific foods, 
preparing them according to traditional methods, and then 
measuring the postprandial blood glucose responses in 
participants after consumption. The findings provided 
insights into dietary recommendations, especially for 
individuals with diabetes or coronary heart disease, by 
identifying foods with lower GI values that are more suitable 
for maintaining stable blood glucose levels. The researchers 
constructed a comprehensive dataset from their own 
experimental data for reference [60], encompassing 
continuous glucose measurements, detailed dietary logs, 
physical activity records, gut microbiota profiles, and various 
blood parameters from an 800‐person cohort. The primary 
purpose of this dataset was to develop and train a machine‐

learning algorithm capable of predicting individualized 
postprandial glycemic responses to different meals. The 
researchers evaluated the Glycemic Index of eight rice 
varieties in Taiwan, including two brown and six white rice 
types in [61]. The dataset was primarily built using both in 
vitro and in vivo methods for different rice varieties. In vitro 
starch digestion tests were conducted to determine the 
predicted Glycemic Index of the rice samples. This approach 
provided valuable insights into the glycemic properties of 
Taiwanese rice varieties, aiding in the development of rice 
with desired health benefits.  

The dataset used in [64] was built by the researchers and 
comprised blood glucose measurements from healthy human 
volunteers who consumed bread samples with varying levels: 
0%, 10%, 15%, and 20% of cassava flour substitution. 
Glucose was used as a reference food to calculate the Glycemic 
Index of each bread variant. Participants consumed the test 
breads after a 10–12‐hour overnight fast, and blood glucose 
levels were recorded at 30‐minute intervals over a 2‐hour 
period following consumption. Then, the GI values of the 
bread samples were determined. The dataset enabled the 
assessment of how substituting wheat flour with cassava 
flour affects postprandial glycemic responses. Also, the study 
found that increasing cassava flour content led to lower 
glycemic responses, with GI values ranging from 91 to 94. The 
authors developed their own dataset for [67] using acoustic 
signals during chewing and swallowing from 50 diabetic 
individuals using an acoustic Micro‐Electro‐Mechanical 
Systems (MEMS) sensor. These signals were then processed 
with a deep learning algorithm to analyze eating patterns and 
formulate a standard procedure aimed at reducing blood 
glucose levels. Reference [74] used a dataset comprising 
clinical measurements from diabetic and non‐diabetic 
individuals by collecting data on Fasting Blood Glucose, 
glycosylated hemoglobin, total plasma homocysteine levels, 
and urinary microalbumin concentrations. Data were 
analyzed using multiple linear regression and machine 
learning algorithms to investigate the relationships between 
glycemic control, hyperhomocysteinemia, and the presence of 
microalbuminuria. The study aimed to identify threshold 
values of FBS and HbA1c associated with microalbuminuria 
and to explore the concurrent association of 
microalbuminuria with hyperhomocysteinemia.  

The dataset in [22] includes comprehensive data on rice 
grain yield and quality traits over a three‐year period, 
spanning 27 different environments in Bangladesh. The 
primary objective was to evaluate the performance of various 
rice genotypes under different environmental conditions to 
identify a stable and adaptive variety with desirable traits, 
including a low Glycemic Index. The data were analyzed using 
statistical methods such as ANOVA and Additive Main effects 
and Multiplicative Interaction analysis to assess genotype‐
environment interactions and stability. This analysis aimed to 
inform breeding programs focused on developing rice 
varieties that are both high‐yielding and suitable for the 
irrigated ecosystems of Bangladesh. The study [20] utilized a 
dataset that comprises continuous glucose monitoring data 
from 102 patients with type 2 diabetes admitted to Cheonan 
Hospital, Soonchunhyang University. This data includes blood 
glucose levels, insulin doses, meal times, and other Electronic 
Medical Records information. The researchers employed a 
multi‐agent reinforcement learning algorithm to perform 
feature selection, aiming to enhance the prediction accuracy 
of adverse glycemic events. The model achieved F1‐scores of 
89.0% for normoglycemia, 60.6% for hypoglycemia, and 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

114 

 

89.8% for hyperglycemia, demonstrating its effectiveness in 
predicting glycemic events. 

Reference [71] utilized retrospective CGM datasets from 
104 individuals who had experienced at least one 
hypoglycemia alert during a three‐day CGM session. These 
datasets were collected from participants who had previously 
undergone CGM monitoring. Among different ML methods 
experimented on the dataset, random forest demonstrated 
the best performance, with an average AUC of 0.966, 
sensitivity of 89.6%, specificity of 91.3%, and an F1 score of 
0.543, concluding that the random forest model holds 
potential for accurately predicting postprandial 
hypoglycemia, which could enhance the effectiveness of 
continuous glucose monitoring and artificial pancreas 
systems. Likewise, reference [39] also follows CMG data from 
124 pregnant women—90 diagnosed with gestational 
diabetes mellitus and 34 healthy controls. This dataset 
included records of 1,489 food intakes, which were 
documented using a mobile application developed for the 
study. The Glycemic Index values for the foods consumed 
were sourced from the University of Sydney’s database and 
incorporated into the app’s food database. The collected data 
were analyzed to develop predictive models for postprandial 
glycemic responses, assessing the impact of incorporating GI 
and GL information on the accuracy of these models.  

Similarly, reference [115] used CGM data from 
individuals, particularly European, Chinese, younger, and 
older participants with Type 1 Diabetes (T1D) Mellitus. 
Benchmark GI values, such as those for glucose and white 
bread, were used as references for comparison and scaling. 
The authors employed machine learning algorithms to 
analyze this data, aiming to predict glycemic levels in real‐
time using constrained Internet of Things devices. The study 
concluded that local, on‐the‐fly forecasting of glycemia is 
feasible with such devices. The dataset used in [78] was built 
by the researchers based on data collected from a clinical trial 
involving 235 participants, including women with gestational 
diabetes mellitus and healthy pregnant women. The data 
includes CGM records, meal‐related information, patient 
characteristics, and survey data. Participants recorded their 
meals in a mobile app while wearing CGM devices to monitor 
blood glucose levels. The dataset captures meal timing, 
composition (e.g., carbohydrate content, GL), and pre‐meal 
CGM trends. Flawed records (e.g., underreported meals) were 
detected and removed to ensure data quality. Models were 
evaluated using cross‐validation and test data from unseen 
participants. Postprandial glucose responses (PPGRs) from 
15 participants who consumed nine standardized meals with 
known macronutrient compositions in [82]. Participants’ 
PPGRs were recorded using continuous glucose monitors 
after they consumed the standardized meals. Each meal’s 
macronutrient content, carbohydrates, proteins, and fats 
were precisely measured. The model’s performance was 
assessed by comparing its macronutrient predictions against 
the actual known values. The proposed sparse coding 
approach consistently outperformed baseline systems based 
on ridge regression and nearest‐neighbors in terms of 
correlation and normalized root mean square error of the 
predictions. This methodology demonstrates the potential of 
using CGM data to automatically estimate dietary intake, 
reducing reliance on self‐reported measures.  

Similarly, reference [107] used a dataset that comprises 
continuous glucose monitoring data collected from 
participants, capturing detailed blood glucose measurements 
over time. The raw CGM data were processed to extract 
relevant features indicative of glycemic patterns, such as 

mean glucose levels, variability metrics, and trends over time. 
Machine learning regression models were trained using the 
extracted features to predict future blood glucose levels. 
These models aimed to forecast glucose trends and potential 
hyperglycemic or hypoglycemic events. This approach 
highlights the effectiveness of using self‐collected CGM 
datasets in developing personalized machine learning models 
for predicting blood glucose levels, which can be instrumental 
in managing diabetes. The dataset in [103] also comprises 
continuous glucose monitoring, which profiles from 20 
healthy subjects who consumed 50 grams of glucose or one of 
four alternative foodstuffs, like chocolate, apple baby food, 
rice squares, or yogurt, at breakfast and dinner over a one‐
week period, resulting in 300 CGM glucose profiles. 
Participants wore CGM devices to continuously record 
interstitial fluid glucose concentrations. They consumed 
specified test foods, each containing 50 grams of 
carbohydrates, at designated meal times, with glucose 
serving as the reference food. The IAUC values obtained from 
the test foods were compared to those from the reference 
food (glucose) to calculate the Glycemic Index for each food 
item. Reference [69] introduced an integrated environment 
for continuous blood glucose monitoring. This system utilizes 
Internet of Things technology to provide real‐time data to 
doctors and caregivers remotely. The researchers developed 
their own dataset by collecting blood glucose mea‐ surements 
using the Freestyle Libre system. This data was then 
transmitted through their IoT‐based platform, enabling 
continuous monitoring and management. The dataset 
facilitated the evaluation of the system’s performance by 
comparing the glucose rates measured with the official 
Freestyle Libre software during the same period. 
Comparably, the dataset in [86] includes continuous glucose 
monitoring readings, records of insulin injections, and 
carbohydrate intake information. The researchers applied 
exponential models to the raw carbohydrate and insulin data 
to simulate absorption processes in the body, aiming to 
enhance the accuracy of their predictive models. By 
incorporating these simulated absorption curves into an RNN 
based on long short‐term memory cells, they sought to 
improve the prediction of future blood glucose levels. 
However, subsequent analysis revealed flaws in the 
experimental techniques, particularly in the model validation 
scheme, which invalidated the reported results and 
conclusions. 

The study [76] employed virtual patient models for T1D 
patients, such as a virtual patient cohort that includes 10 
adults and 10 adolescents. These models simulate various 
physiological responses to insulin treatment, allowing for 
controlled experimentation without human participants. In‐
Silico Data: The virtual patient models were used to simulate 
and evaluate the performance of the machine learning‐based 
artificial pancreas algorithm under various scenarios, 
providing preliminary insights into its potential effectiveness 
and safety. In‐Vivo Data: The clinical trials with human 
participants were conducted to validate the algorithm’s 
performance in real‐world settings, assessing outcomes such 
as time‐in‐range (TIR), hypoglycemic events, and overall 
glucose control. The study [18] creates a dataset by utilizing 
details of 12 healthy volunteers (6 men and 6 women) aged 
between 20 and 30 years. Four traditional Omani rice dishes 
were selected: white rice, biryani, kabsa, and maqboos. 
Glucose was used as the reference food for determining the 
Glycemic Index. Blood glucose levels were measured at 
intervals of 15, 30, 45, 60, 90, and 120 minutes after 
consumption. With the data, the GI and GL were calculated. 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

115 

 

The study found that white rice had the highest GI value 
(77.3), while the other rice dishes had moderate GI values. 
This information is valuable for dietary planning, especially 
for individuals managing blood sugar levels. The dataset in 
[19] comprised 20 infant cereal prototypes, each with varying 
macronutrient compositions, particularly in glycemic 
carbohydrates (ranging from 51 to 76 grams per 100 grams), 
and was then utilized to validate a predictive model for 
estimating glycemic responses based on nutritional 
composition. The data were collected through four 
independent studies measuring the postprandial glucose 
responses of these cereal prototypes in healthy adults. The 
collected data were then applied to a predictive model 
previously developed to estimate the GI and GL of breakfast 
cereals based on their macronutrient composition. This 
model quantifies both the impact of glycemic carbohydrates 
and the GI‐lowering effects of other macronutrients such as 
proteins, fats, and fibers. 35 adults with prediabetes (17 
females, 18 males; mean age 54.2 years; mean BMI 32.44 
kg/m²) have contributed data to [32]. Participants 
underwent a controlled feeding study, consuming either a low 
Glycemic Index (LGI) diet (GI < 35) or a High Glycemic Index 
(HGI) diet (GI > 70) for four weeks, following a two‐week 
control diet (GI = 55–58). The researchers conducted 4‐hour 
meal tolerance tests to assess insulin sensitivity, insulin 
secretion, and β‐cell function. The collected data were 
analyzed using mathematical modeling to evaluate the impact 
of dietary Glycemic Index on β‐cell function in individuals 
with prediabetes.  

The 1993 study titled “Prediction of Glycemic Index for 
starchy foods”, reference [113] analyzed 18 starchy foods to 
examine the relationship between their Glycemic Index and 
chemical components such as protein, fat, phytic acid, and 
Total Dietary Fiber. For each food item, the dataset included 
measurements of protein, fat, phytic acid, and TDF present in 
portions containing 50 grams of available carbohydrate. The 
researchers employed regression analysis to explore 
associations between the GI and the chemical components of 
the foods. They found significant correlations (P < 0.05) 
between GI and TDF, protein, and phytate, also the analysis 
suggested that the method of food preparation and the 
characteristics of starch and starch granules might be more 
critical in predicting GI among starchy foods than the content 
of any single component. Reference [108] involved a dataset 
that comprised glycaemic index measurements from a 
randomized crossover meal test with 28 healthy young men. 
Participants consumed 13 different breakfast meals and a 
reference meal, each containing 50 grams of available 
carbohydrates but varying significantly in energy and 
macronutrient composition. Venous blood samples were 
collected over a two‐hour period to analyze glucose and 
insulin responses. The study aimed to assess whether the GI 
of mixed meals, calculated using standard GI tables, 
accurately predicted the measured GI.  

The dataset in [104] encompasses various cereal and 
legume‐based food products, with detailed information on 
their macronutrient compositions, including carbohydrate, 
protein, fat, and fiber contents. Additionally, the dataset 
includes measured Glycemic Index values for these foods, 
obtained through in vivo testing. Researchers then developed 
predictive models to estimate the GI of foods based on their 
macronutrient profiles. These models aimed to identify 
relationships between macronutrient composition and GI, 
facilitating the prediction of GI for similar foods without the 
need for extensive in vivo testing. Reference [106] introduced 
a mixed‐integer programming model aimed at minimizing the 

total daily glycemic load of foods while satisfying daily 
nutritional and serving size requirements. The dataset 
employed comprises 177 foods, with their nutritional 
information and GL values sourced from the U.S. Department 
of Health and Human Services and the U.S. Department of 
Agriculture (USDA) guidelines. This dataset is not a standard 
benchmark but is constructed by the researchers using 
publicly available nutritional data. In the study, the dataset is 
utilized to perform experimental analyses, applying robust 
optimization techniques to account for uncertainties in GL 
values. Participants for the research [90] were 79 children 
diagnosed with type 1 diabetes. For each participant, various 
factors were recorded, including demographic information, 
biological markers, and socioeconomic status. Machine 
learning algorithms were employed to train predictive 
models using the selected features. The objective was to 
forecast glycemic control, focusing on achieving an A1C level 
below 7.5%, as recommended by organizations such as the 
American Diabetes Association (ADA) and the International 
Society for Pediatric and Adolescent Diabetes (ISPAD). This 
approach aimed to enhance the understanding of factors 
influencing glycemic control and to improve predictive 
capabilities in clinical settings, ultimately contributing to 
better management strategies for children with type 1 
diabetes. Reference [31] comprised dietary records from 131 
participants following various modern diets. Participants’ 
dietary intakes were recorded and analyzed using the 
Nutrition Data Systems for Research (NDSR) software. This 
process involved detailed logging of food consumption to 
assess diet quality. The collected data were used to calculate 
three key dietary indices: Healthy Eating Index, GI, and GL, 
which evaluate diet quality based on adherence to dietary 
guidelines. Artificial Intelligence and Machine Learning 
techniques were applied to the dataset to identify predictors 
of the dietary indices. Factors such as whole fruit and whole 
grain consumption were found to be significant predictors of 
HEI, while carbohydrate intake was a common predictor for 
both GI and GL. The dataset comprises data from laboratory 
analyses conducted on wheat‐based bread samples fortified 
with varying concentrations (0.25% to 2%) of Euryale ferox 
seed shell extract in [26]. The researchers assessed the 
inhibitory effects of EFSSE on α‐amylase and α‐glucosidase 
activities, and evaluated the in vitro starch digestibility 
(IVSD) and predicted Glycemic Index of the bread samples. 
Advanced computational techniques, including Swarm 
Intelligence supervised neural network modeling, were 
employed to simulate digestion kinetics and predict the 
Glycemic Index, providing insights into the potential of EFSSE 
as a functional additive for producing lower Glycemic Index 
bread.  

The study titled “Moroccan Food Dataset for Food Image 
Recognition Towards Glycemic Index Estimation”, [11] 
introduced the MFOOD‐70 dataset, a collection of 70 
Moroccan food categories comprising 14,000 images. This 
dataset was specifically developed by the authors to enhance 
food image recognition and facilitate Glycemic Index 
estimation. The images were sourced from web scraping and 
existing datasets, ensuring a diverse representation of 
Moroccan cuisine. The dataset was utilized to train and 
evaluate convolutional neural network models, aiming to 
improve the accuracy of food recognition systems and 
support dietary monitoring applications. In reference [28], 
the dataset comprises data from 10 healthy non‐diabetic 
volunteers (5 males and 5 females). Each participant 
consumed 50 grams of carbohydrate from different black rice 
cultivars after an overnight fast. Blood glucose levels were 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

116 

 

measured at intervals of 0, 15, 30, 45, 60, 90, and 120 minutes 
post‐consumption. The researchers calculated the 
Incremental IAUC for each rice cultivar and compared it to the 
IAUC of a reference food (glucose) to determine the Glycemic 
Index values. The study found that the GI values of the black 
rice cultivars ranged from 44.6 to 59.7, indicating that these 
cultivars have a low to medium GI. The dataset was built by 
the researchers specifically for this study to assess the GI of 
selected black rice cultivars in Bangladesh. The dataset 
described in [24] primarily consists of various rice genotypes 
used to study their Glycemic Index and associated 
biochemical properties such as resistant starch content, 
amylose content, and other indicators of starch digestibility. 
These genotypes were sourced from different ecologies or 
traits and analyzed using in vitro methods to determine their 
GI values and RS levels. This dataset aids in understanding 
how genetic and biochemical variations in rice influence its 
digestion and postprandial glucose response, with 
implications for dietary recommendations and crop 
improvement strategies. Reference [23] developed a custom 
dataset specifically by formulating various bread samples by 
substituting wheat flour with chickpea flour, red chicory 
powder, and different types of resistant starch. They then 
assessed the predicted Glycemic Index and technological 
properties of these bread formulations. The dataset includes 
measurements of pGI, moisture content, volume, specific 
volume, baking loss, and texture parameters such as 
hardness, cohesiveness, and chewiness. This comprehensive 
dataset enabled the researchers to analyze how each 
ingredient influenced the bread’s Glycemic Index and 
technological characteristics. Reference [65] encompasses 
various food items, each characterized by both nutrient‐
based information and non‐nutrient physicochemical 
properties such as texture, water activity, glycemic potential, 
satiety potential, and shelf life. The researchers employed 
data mining techniques to analyze the compiled dataset, 
aiming to establish correlations between the degree of food 
processing and the physicochemical properties of the foods. 
By examining these relationships, they sought to develop a 
comprehensive quantitative index that reflects the extent of 
food processing, moving beyond traditional nutrient‐based 
indices. This holistic index is intended to provide a more 
nuanced understanding of how processing affects food 
quality and health implications. 106 participants, including 
53 colorectal cancer cases and 53 family members from 
diverse ethnic backgrounds, participated in creating the 
dataset of reference [66]. The data encompassed individual 
dietary parameters, health outcomes, and demographic 
information.    The researchers employed machine learning 
validation procedures, such as the ensemble method and 
generalized regression prediction, to analyze the data. 
Significant dietary predictors identified included whole fruit, 
milk or milk alternatives, whole grains, saturated fat, and oils 
and nuts. These findings highlight the importance of specific 
dietary components in promoting healthy eating habits 
among multi‐ethnic colorectal cancer families. The dataset 
employed in [36] comprises various Sri Lankan starchy 
tubers, including arrowroot, cassava, potato, purple yam, 
sweet potato, and white yam. The researchers collected these 
tubers from local sources and prepared them under 
controlled laboratory conditions to assess their starch 
hydrolysis indices. Each tuber was subjected to enzymatic 
digestion to measure the rate and extent of starch breakdown 
over time. These measurements enabled the calculation of the 
Hydrolysis Index for each tuber, which serves as an indicator 
of the potential glycemic response upon consumption. The 

findings provide valuable insights into the nutritional 
properties of these traditional Sri Lankan tubers, particularly 
concerning their impact on blood sugar levels.  

The study [29] utilizes a dataset comprising real‐world 
data from individuals with Type 1 Diabetes. This dataset 
includes patient‐specific information such as blood glucose 
levels, insulin doses, and nutritional intake. The researchers 
collected this data to develop a machine learning model 
capable of predicting postprandial blood glucose levels at 
various time intervals (15, 30, 45, and 60 minutes) following 
a meal. By incorporating these nutritional factors, the model 
aims to enhance the accuracy of blood glucose predictions, 
thereby supporting better management of T1D. AI4FoodDB is 
a public database developed by researchers to support 
personalized e‐health nutrition and lifestyle studies in [92]. It 
was constructed from a nutritional weight loss intervention 
involving 100 overweight and obese participants over one 
month. The dataset includes various types of data collected 
through manual methods, clinical assessments, and digital 
tools such as wearable devices. The database comprises 
several distinct datasets: anthropometric measurements, 
lifestyle and health, nutrition, biomarkers, physical activity, 
sleep activity, emotional state, etc. These datasets are utilized 
to analyze the relationships between various lifestyle, 
biological, and digital factors and health outcomes. By 
integrating diverse data sources, AI4FoodDB facilitates the 
development of artificial intelligence techniques aimed at 
advancing personalized healthcare.  

In the study  [93] the researchers constructed a custom 
fruit dataset specifically for their research on diabetic 
patients’ daily diets. This dataset was not sourced from 
existing benchmarks but was developed to facilitate the 
identification of fruits with high sweetness and low glycemic 
load values. The dataset was utilized to train and evaluate an 
improved Faster R‐CNN network, which incorporated an 
attention mechanism module during feature extraction, 
adjusted the anchor aspect ratio of the Region Proposal 
Network (RPN), and implemented a fusion update operation 
in the fully connected layer. These enhancements aimed to 
improve the precision and recall rates of fruit recognition, 
ultimately assisting diabetic patients in making informed 
dietary choices. 

These trials provided empirical evidence supporting the 
algorithm’s efficacy and safety in managing T1D. The authors 
collected spectral data for the study [77] from rice samples 
using a portable Near-Infrared sensor operating in the 740–
1070 nm wavelength range. The collected spectral data were 
then analyzed using machine learning techniques, including 
principal component analysis, linear discriminant analysis, 
random forest classifier, and partial least squares regression, 
to develop predictive models for rice quality attributes such 
as Glycemic Index, amylose content, and viscoelasticity. These 
models aimed to provide rapid, on‐site evaluation of rice 
quality. Similarly, a dataset collected by the researchers 
through a randomized trial involving adults with prediabetes 
using waist‐worn or wrist‐worn wearables to monitor their 
activity patterns in reference [88]. Baseline information, 
including demographics, medical history, and laboratory test 
results, was also gathered. The study developed predictive 
models to assess changes in hemoglobin A1c levels, an 
indicator of glycemic control. The models compared 
traditional regression methods with machine learning 
approaches, finding that ensemble machine learning methods 
provided better predictions. Additionally, incorporating 
wearable data alongside baseline information improved 
prediction accuracy. Notably, wrist‐worn wearables yielded 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

117 

 

more accurate predictions compared to waist‐worn devices. 
These findings suggest that integrating wearable‐derived 
activity data with clinical information can enhance the 
prediction of glycemic control changes in individuals with 
prediabetes.  

Another research focuses on developing a personal 
device to assist diabetic patients in managing insulin therapy 
[105]. The researchers collected data from diabetic patients, 
including blood glucose levels, insulin dosages, meal 
information, and other relevant health metrics. This data was 
gathered using the developed personal device integrated with 
various sensors and input methods. The collected data was 
utilized to test and refine the device’s algorithms for 
calculating insulin dosages, considering factors such as 
patient weight, glucose levels, physician recommendations, 
and carbohydrate absorption. The dataset supported the 
integration of the personal device with a Glycemic Index 
information system, nurses’ and physicians’ desktop 
applications, and a patient web portal, facilitating 
comprehensive diabetes management within an Ambient 
Assisted Living environment. Reference [79] comprised 92 
blood samples collected from individuals in Santa Cruz do Sul, 
Brazil, with informed consent. The dataset contains mid‐
infrared spectra of peripheral blood samples, which were 
analyzed using diffuse reflectance infrared spectroscopy 
(DRIFTS). The spectral data were processed to quantify 
biochemical parameters such as total cholesterol, using 
artificial neural networks. The ANN achieved a correlation 
coefficient (r) of 0.81 and a root mean square error (RMSE) of 
30.14 in the preliminary trial. Future plans include expanding 
the dataset to 500 samples to enhance accuracy and include 
other parameters like HDL, LDL, triglycerides, and glucose. 13 
qualified individuals (8 men and 5 women) participated in 
[81]. Blood samples were taken in the fasting state and at 15, 
30, 45, 60, 90, and 120 minutes after ingestion. The blood 
glucose levels measured at the specified intervals were used 
to assess the body’s glycemic response to both the reference 
and test foods. Results suggest that the nutritional product 
elicits a lower glycemic response compared to glucose, 
making it suitable for individuals managing diabetes mellitus. 
Reference [38] utilized a dataset collected by 18 healthy 
volunteers with fasting plasma glucose levels below 100 
mg/dL. Participants consumed three different test foods in a 
randomized sequence: a complete nutrition drink containing 
retrograded starch, a glucose solution, and white bread. 
Plasma glucose and insulin levels were measured at baseline 
and at multiple time points up to 180 minutes post‐
consumption. The dataset facilitated the assessment of 
postprandial insulin responses, revealing that the complete 
nutrition drink led to a sustained increase in plasma insulin 
levels over the 3‐hour period, in contrast to the more rapid 
decline observed with glucose solution and white bread.  

The dataset in [89] collected as part of the “Smart District 
4.0 Project”, supported by the Italian Ministry of Economic 
Development. The study involved six patients with diabetes. 
Glycemic values were recorded every 3 minutes using 
specialized monitoring devices. The number of observations 
varied among patients; for instance, Patient A had 243 
observations, while Patient B had approximately 13,204 
observations. Eight different algorithms were employed to 
predict the glycemic status of the patients: artificial neural 
network, Probabilistic Neural Network, Polynomial 
Regression, Gradient Boosted Trees Regression, Random 
Forest Regression, Simple Regression Tree, Tree Ensemble 
Regression, and Linear Regression. The models were 
evaluated based on their ability to minimize four statistical 

errors: Mean Absolute Error (MAE), Mean Squared Error 
(MSE), Root Mean Squared Error, and Mean Signed Difference 
(MSD). The study aimed to identify the most efficient 
algorithm for predicting glycemic status by comparing these 
errors across the different models. Twenty-six participants, 
who contributed to the creation of the dataset of [87], also 
wore non‐invasive, wrist‐worn wearable devices in 
conjunction with continuous glucose monitors for 8–10 days 
after undergoing a clinical HbA1c measurement. The 
wearables captured physiological data, which were then 
analyzed to estimate glucose variability metrics and HbA1c 
levels. The study developed 27 models to estimate glucose 
variability metrics using data from the non‐invasive 
wearables, achieving high accuracy (mean average percent 
error (MAPE), of less than 10%) in 11 of these models. 
Additionally, the HbA1c estimation model achieved a MAPE of 
5.1% on an external validation cohort. This proof‐of‐concept 
study demonstrated the feasibility of using non‐invasive 
wearables for glycemic monitoring, potentially offering a 
more convenient and less invasive method for patients to 
monitor their glucose levels and HbA1c remotely. 

Likewise, the dataset in [35] comprises data collected 
from participants using wearable sensors and mobile devices 
to monitor food intake, physical activity, and corresponding 
blood glucose levels. Relevant features, such as meal timing, 
nutritional content, activity type, duration, and intensity, are 
extracted from the raw data to serve as inputs for the 
predictive models. Deep Learning algorithms are employed to 
analyze the extracted features and predict blood glucose 
levels based on observed patterns in food consumption and 
physical activity. This approach aims to develop a non‐
invasive method for monitoring blood glucose levels by 
leveraging deep learning techniques to interpret lifestyle 
data. 1,159 adults aged 20‐74 years with type 2 diabetes and 
HbA1c levels between 6.0‐8.9 (42‐74 mmol/mol) were 
involved in [100]. The participants’ health metrics were 
continuously monitored using IoT devices, and this real‐time 
data was analyzed to assess the effectiveness of the IoT‐based 
intervention on glycemic control over a 52‐week period. The 
primary endpoint was the change in HbA1c levels from 
baseline to the final measurement at 52 weeks. The study 
concluded that the IoT‐based approach did not significantly 
reduce HbA1c in patients with type 2 diabetes. The authors 
suggested that incorporating daily glycemic control data and 
HbA1c levels into the IoT‐based intervention may be 
necessary to improve glycemic control. 

5.2 Utilization of benchmark datasets in technological 
approaches for Glycemic index assessment 
The 2016 study [62] utilized datasets from two 

European Union‐funded projects: 
• DIAdvisor (EU FP7‐funded project): This dataset 

comprised clinical trial data, including intermittent blood 
glucose measurements from patients with type 1 diabetes. 
The data were used to calibrate and test nocturnal 
hypoglycemia (NH) predictors based on various glycemic 
control indices (GCI). 

• AMMODIT (EU Horizon 2020‐funded project): This dataset 
was employed to validate the portability and effectiveness 
of the proposed NH prediction approach across different 
patient populations. 

The authors developed a method to predict nocturnal 
hypoglycemia by aggregating predictors constructed from 
different GCIs, such as the Low Blood Glucose Index. They 
applied machine learning techniques to combine these 
predictors, aiming to enhance the accuracy of NH predictions. 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

118 

 

The datasets provided the necessary blood glucose 
measurements to calibrate and test the performance of the 
aggregated predictors, demonstrating improved sensitivity 
and specificity in predicting NH events. The dataset used in 
[63] includes Glycaemic Index values from an existing 
comprehensive list (e.g., a glycaemic index corpus) and food 
nutritional composition data (such as macronutrient content) 
from organizations like the USDA. The dataset is not created 
by the authors but sourced from these established 
repositories. The study highlights challenges in integrating 
these datasets due to differences in how foods are labeled and 
categorized. The data is used to predict the GI of foods based 
on their biochemical properties using machine learning 
techniques. The authors manually cross‐linked a subset of 
100 food entries to ensure reliability and used features like 
water, energy, protein, carbohydrate, sugar, fiber, and lipid 
content to build predictive models. This approach aimed to 
explore the feasibility of using widely available food data for 
practical GI prediction.  
 Reference [72] exploits a dataset of 55, 102 cooking 
recipes collected from the Allrecipes website. This dataset 
was built by the researchers specifically for their analysis. A 
subset of 1,000 recipes was selected for further analysis. 
These recipes were annotated for glycemic impact through 
online crowdsourcing using Amazon Mechanical Turk. From 
the collected data, both textual features (e.g., ingredients and 
cooking directions) and nutritional features (e.g., 
carbohydrate and sugar content) were extracted to represent 
each recipe. Machine Learning models were trained using the 
annotated subset to classify recipes as healthy or unhealthy 
for diabetics based on their glycemic impact. This approach 
combines online crowdsourcing and machine learning to 
estimate the glycemic impact of cooking recipes, offering a 
data‐driven method to assist diabetics and pre‐diabetics in 
making informed dietary choices. Reference [17] presented a 
systematic six‐step methodology to assign Glycemic Index 
values to over 600 foods in the Venezuelan food composition 
database. The process begins with compilation, where GI 
values from international sources are gathered. Next, the 
matching step aligns these values with local foods based on 
ingredient and preparation similarities. For foods lacking 
direct matches, estimation is employed by analyzing 
macronutrient composition and comparing it with analogous 
foods. This is followed by validation, ensuring accuracy 
through cross-referencing with literature and expert 
opinions. The validated GI values are then integrated into the 
national database. Finally, documentation provides a 
transparent record of the sources and rationale, ensuring 
traceability and facilitating updates. This methodology 
enriches the database with reliable GI values, supporting 
dietary planning and nutritional evaluation tailored to the 
Venezuelan context.  

In 2019, reference [33] utilized a dataset comprising the 
macronutrient compositions of 42 breakfast cereals. This 
dataset was collected from existing nutritional information 
available for these products. The researchers did not generate 
new experimental data but instead relied on published 
macronutrient profiles to develop their predictive model. The 
dataset was employed to create a model that predicts the 
Glycemic Index and glycemic load of foods based on their 
macronutrient content. By analyzing the relationship 
between the macronutrient composition and the GI/GL 
values, the model quantifies the impact of glycemic 
carbohydrates and the GI‐lowering effects of other nutrients 
such as proteins, fats, and fibers. Reference [30] utilizes the 
PIMA Indians Diabetes dataset, a benchmark dataset 

provided by the National Institute of Diabetes and Digestive 
and Kidney Diseases. The numerical data from the PIMA 
dataset are transformed into image representations. Each 
feature is assigned a specific location and size within the 
image based on its importance, determined using the ReliefF 
feature selection algorithm. This approach enables the 
application of convolutional neural networks designed for 
image data. To enhance the dataset, data augmentation 
techniques are applied to the generated images, artificially 
increasing the number of training samples and improving 
model robustness. The augmented image data are used to 
train deep learning models, specifically ResNet18 and 
ResNet50 CNN architectures, for predicting pre‐diabetic 
conditions. The innovative approach leverages image‐based 
deep learning techniques to enhance the prediction of pre‐
diabetic conditions using a well‐established benchmark 
dataset. Reference [91] employed data from the All of Us 
Research Program, a comprehensive initiative by the U.S. 
National Institutes of Health aimed at gathering health data 
from diverse populations to advance precision medicine. 
Researchers applied various machine learning algorithms, 
including random forest, extreme gradient boost, logistic 
regression, and a weighted ensemble model, to predict 
uncontrolled diabetes. They identified patients aged 18 and 
above with diabetes from the All of us dataset and defined 
uncontrolled diabetes based on specific International 
Classification of Diseases codes. The models incorporated 
features such as basic demographics, biomarkers, and 
hematological indices. Among these, the random forest model 
demonstrated the highest performance, achieving an 
accuracy of 80% and an area under the receiver operating 
characteristic curve of 0.77. Key predictors of uncontrolled 
diabetes included serum potassium levels, body weight, 
aspartate aminotransferase, height, and heart rate.  

Reference [98] involved the creation of a 
comprehensive database by the researchers themselves. This 
database integrates Glycemic Index and glycemic load values 
with dietary data from the National Health and Nutrition 
Examination Survey (NHANES) spanning 1999 to 2018. The 
researchers employed an artificial intelligence‐enabled 
model to align GI values from existing databases with 
NHANES food codes. This process was manually validated to 
ensure accuracy, resulting in GI values covering 99.9% of total 
carbohydrate intake. This newly developed database serves 
as a valuable resource for large‐scale epidemiologic studies, 
enabling researchers to assess the impact of carbohydrate 
quality on health outcomes within the U.S. population. 
Reference [99] comprises 1,000 records from the Diabetes 
Complication Early Warning Dataset provided by the National 
Clinical Medical Sciences Data Center. The dataset underwent 
preprocessing to address missing values and outliers. Feature 
selection was performed using information gain to identify 
the most relevant variables. Subsequently, the authors 
developed a diabetic retinopathy risk prediction model 
employing the CatBoost algorithm, an advanced machine 
learning technique. To enhance the interpretability of the 
model’s predictions, they applied SHAP values, which 
elucidate the contribution of each feature to the model’s 
output. This approach enabled the identification of key risk 
factors associated with diabetic retinopathy, such as poor 
renal function, elevated blood glucose levels, liver disease, 
hematonosis, and dysarteriotony. The integration of Machine 
Learning with interpretable models facilitated a more 
transparent understanding of the factors influencing diabetic 
retinopathy risk. PhysioNet, a public database, is used in [21]. 
This dataset comprises physiological and nutritional 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

119 

 

information collected via wearable devices and dietary 
surveys. The researchers employed this existing dataset to 
analyze the relationships between blood glucose levels and 
various physiological and nutritional factors. They conducted 
correlation analyses, multiple regression analyses, and one‐
way analyses of variance to explore how different 
physiological indicators and nutritional values are associated 
with blood glucose fluctuations. 

The public dataset, D1NAMO is used in [83], an open‐
source collection of real‐time Glycemic Index readings. This 
dataset comprises continuous glucose monitoring data 
collected every 5 minutes from 9 diabetic patients and six 
times daily from 20 non‐diabetic individuals. The data was 
gathered under normal conditions using the Zephir 
Bioharness 3 wearable device. They implemented an auto‐
adaptive algorithm for optimizing ARIMA model parameters, 
enabling real‐time predictions in an online learning 
environment. The dataset’s comprehensive CGM readings 
facilitated the training and evaluation of these models, aiming 
to enhance glycemic control through accurate short‐term 
forecasts. Similarly, reference [73] utilizes benchmark 
datasets from continuous glucose monitoring devices worn 
by patients with Type 1 Diabetes. The CGM data is segmented 
into feature vectors using a sliding window technique, 
capturing blood glucose readings over specific time intervals. 
This method generates training examples that reflect the 
temporal dynamics of glucose levels. Machine learning 
models, including Support Vector Regression and Multilayer 
Perceptron, are trained on the processed and balanced 
datasets. 

The article [40] is a consensus statement that reviews 
existing research and provides expert opinions on the 
Glycemic Index, glycemic load, and glycemic response. As a 
consensus statement, it does not introduce new experimental 
data or utilize a specific dataset. Instead, it synthesizes 
findings from numerous studies to offer guidance on the 
application and interpretation of GI and GL in nutrition 
science and public health. The document serves to 
consolidate scientific understanding and provide 
recommendations based on a comprehensive review of 
existing literature. The article [27] is a review that 
synthesizes existing research on the Glycemic Index of rice 
and its products. It does not introduce a new dataset; rather, 
it compiles and analyzes data from various studies published 
up to December 2022. The authors conducted a 
comprehensive literature review, gathering information from 
the Web of Science and Scopus databases. They categorized 
the findings into four main sections: basic information about 
starch digestion and recent advanced measurement methods, 
the mechanism of the effect of various factors on GI, recent 
advanced technologies to modulate GI, and a table of the 
Glycemic Index for rice and rice products in different 
countries. This compilation provides an overview of the GI 
values of different rice varieties and discusses the impact of 
various factors and processing techniques on the GI of rice 
products. 

The review paper [94] provides consensus guidelines for 
machine learning practitioners in diabetes care. It reviews 
common features used in machine learning applications for 
glucose control and offers an open‐source library of functions 
for calculating these features. Additionally, it provides a 
framework for specifying datasets using data sheets and 
reviews current datasets available for training algorithms, 
along with an online repository of data sources. These 
resources are designed to improve the performance and 
translatability of new machine learning algorithms developed 

in the field of diabetes. Reference [95] is also a systematic 
review that evaluates various mobile computer vision‐based 
approaches for food classification, volume estimation, and 
nutrient estimation. As a review, it does not introduce a new 
dataset but rather examines existing methods and the 
datasets they utilize. The datasets referenced in the reviewed 
studies vary; some are proprietary datasets developed by 
researchers, while others are benchmark datasets commonly 
used in the field. These datasets are employed to train and 
validate models that can accurately classify food items, 
estimate portion sizes, and assess nutritional content, which 
are crucial for managing dietary intake in individuals with 
diabetes. The review article [101] published in the Alexandria 
Engineering Journal in 2024 conducted a comprehensive 
literature review, synthesizing information from various 
studies and sources to discuss the advancements and trends 
in wearable glucose monitoring technologies. This approach 
involves aggregating and analyzing existing research findings 
rather than applying a new or benchmark dataset. 

The dataset used in reference [28] comprises two parts: 
The “Indian Images Top (20)” dataset available on Kaggle, 
containing 3996 images from 20 different Indian food classes, 
and a custom dataset created for the research, which includes 
nutritional information like Glycemic Index, protein, fats, and 
carbohydrates for the food items. The Kaggle dataset serves 
as the primary image dataset for training and testing the 
Inception V3 model for food classification. Data augmentation 
techniques such as rotation, shearing, and horizontal flipping 
were applied to increase the dataset to 4996 images, ensuring 
better model generalization. The custom dataset 
complements this by providing essential nutritional details to 
enable personalized food recommendations. Thus, the 
research combines a public benchmark dataset with a 
custom‐built dataset tailored for its objectives. The dataset in 
the document “Index of foods: A review [34] is derived from 
existing literature, using databases like MEDLINE, PubMed, 
Scielo, and Google Scholar. It is not an original dataset created 
by the authors but rather a compilation of data from prior 
research studies. This approach allows the authors to 
summarize findings on the Glycemic Index and its influence 
on health, using these benchmark sources. The dataset is 
utilized to analyze patterns and outcomes related to dietary 
habits, carbohydrate types, and their metabolic impacts, 
emphasizing their relevance in managing chronic diseases 
like diabetes and obesity. 

5.3 Studies utilizing both benchmark and custom data 
The dataset used in [44] combines experimental data 

and benchmark GI data produced specifically for this 
research. Experimental data were measured through in vivo 
methods based on FAO/WHO protocols. These 
measurements are combined with benchmark GI data 
(sourced from international tables and scientific literature) to 
train an artificial neural network. These data are then 
analyzed using an ANN to establish a predictive model for 
Glycemic Index values. This approach provides a more cost-
effective and faster alternative to in vivo testing, allowing for 
the prediction of GI with high accuracy, as evidenced by the 
study’s cross‐validation results (R² = 0.89). Another study has 
included [43] both benchmark values for Glycemic Index from 
prior in vivo studies and new experimental data derived from 
in vitro digestion methods combined with High‐Performance 
Liquid Chromatography (HPLC) analysis. Also, in reference 
[42], the dataset used is drawn from both experimental data 
and existing GI benchmarks. It includes measured Glycemic 
Index values for foods like rice and breakfast cereals, using 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

120 

 

glucose and starchy reference foods. These were collected 
through controlled experiments involving European, Chinese, 
younger, and older participants. Benchmark GI values, such as 
those for glucose and white bread, were used as references 
for comparison and scaling. The dataset was employed to 
explore how reference foods influence observed GI values 
across populations. It was also used to assess variability in GI 
results due to participant factors like ethnicity or age, and to 
evaluate the appropriateness of different reference foods for 
GI testing.  

The review paper [37] presents an updated compilation 
of Glycemic Index and GL values for various foods. The dataset 
used in this study is an extensive collection of both published 
and unpublished GI values sourced from global research 
conducted between January 1, 2008, and June 30, 2020. The 
authors systematically reviewed and tabulated these sources, 
adhering to the International Standards Organization (ISO) 
methodology to ensure data reliability. The dataset includes 
over 4,000 food items, representing a 61% increase from the 
previous edition published in 2008. This comprehensive 
dataset serves as a valuable resource for researchers and 
healthcare professionals, facilitating a better understanding 
of the glycemic impact of various foods and aiding in the 
development of dietary recommendations. Reference [97] 
employed datasets from two primary sources: in‐silico 
simulations and the OhioT1DM dataset. The in‐silico cohorts 
comprised 20 and 47 virtual patients, respectively, designed 
to mimic real‐world scenarios. The OhioT1DM dataset is a 
publicly available collection of data from individuals with 
type 1 diabetes, including continuous glucose monitoring 
data. The researchers employed a heterogeneous ensemble 
method combining artificial neural networks, random forests, 
and logistic regression to develop a meal detection model. 
This model was trained and tested on both the in‐silico and 
OhioT1DM datasets to enhance its robustness and accuracy. 
The ensemble majority voting approach achieved high 
sensitivity and precision in detecting unannounced meals, 
thereby improving postprandial glucose control.  

A review paper [37] presents an updated compilation of 
Glycemic Index and glycemic load values for various foods. 
The dataset used in this study is an extensive collection of 
both published and unpublished GI values sourced from 
global research conducted between January 1, 2008, and June 
30, 2020. The authors systematically reviewed and tabulated 
these sources, adhering to the International Standards 
Organization methodology to ensure data reliability. The 
dataset includes over 4,000 food items, representing a 61% 
increase from the previous edition published in 2008. This 
comprehensive dataset serves as a valuable resource for 
researchers and healthcare professionals, facilitating a better 
understanding of the glycemic impact of various foods and 
aiding in the development of dietary recommendations. 

6. Research gaps and future directions 

The timeline analysis (Figure 7) reveals that a significant 

portion of Glycemic Index research was conducted during the 

2015‐2024 period, reflecting a concentrated effort to explore 

foundational techniques in GI prediction and analysis. 

However, much of this work has relied on traditional 

statistical methods and machine learning, with limited 

integration of emerging technologies such as Explainable AI, 

Deep Learning, and Reinforcement Learning. The consistent 

presence of machine learning (green category) underscores 

its foundational role in GI research, but future studies should 

explore ensemble methods and meta‐learning to improve 

predictive performance, particularly for diverse food types 

and ripeness stages. Deep learning’s growing role since 2015 

presents opportunities for leveraging advanced architectures 

such as transformer‐based models or multimodal learning. By 

combining diverse data sources, including biochemical food 

properties, imaging data, and textual descriptions, these 

methods can significantly enhance the accuracy and scope of 

GI prediction. Statistical learning, while enduringly relevant, 

can be effectively combined with modern deep learning 

approaches to create hybrid models for interpretable and 

robust predictions. Reinforcement Learning, with only one 

study in the current collection, is an underexplored yet 

promising direction. RL agents can dynamically predict GI 

based on ripeness levels by interacting with sensor networks 

in food supply chains, using iterative feedback to enhance 

accuracy. Another promising application involves RL‐driven 

dietary recommendation systems that adjust in real‐time 

based on users’ blood glucose levels, dietary preferences, and 

lifestyle factors such as exercise and stress. With 

advancements in wearable technology and IoT devices, RL‐

based systems could revolutionize personalized glycemic 

management by minimizing glycemic spikes and optimizing 

dietary plans tailored to individual needs. 

The mapping of the literature by region revealed notable 
insights into the geographical distribution of research 
activities. As shown in Figure 9, the majority of GI‐based 
research utilizing AI technologies has been conducted in 
North American countries. Additionally, several Asian 
countries, such as India, have contributed significantly to this 
field with a substantial number of publications. These 
findings highlight the global interest in GI‐based research 
while also emphasizing regional disparities in research 
output, suggesting opportunities for further contributions 
from underrepresented regions. 

 
Figure 9: Number of papers occurrences by geographic region 

The dataset analysis (Figure 8) highlights the increasing 
use of custom datasets, which often integrate continuous 
glucose monitoring (CGM) data, wearables, and IoT devices to 
monitor food intake, activity patterns, and health metrics. 
While custom datasets enable granular insights, existing 
benchmark datasets like NHANES, DIAdvisor, AMMODIT, 
D1NAMO, and OhioT1DM remain crucial for stan‐ 
standardization. Hybrid datasets that combine custom and 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

121 

 

benchmark data are emerging as a powerful approach, 
offering the reliability of standardized data along with the 
specificity of real‐world measurements. Additionally, 
datasets sourced from recipe platforms like allrecipes 
provide valuable nutritional insights, complementing 
structured survey data and enhancing the 
comprehensiveness of research. 

Looking ahead, integrating these diverse datasets with 
AI‐driven models holds immense potential for advancing GI 
research. Automated image analysis can improve ripeness 
and portion-size estimation, while wearable data and IoT 
systems can facilitate real‐time and longitudinal studies. Such 
approaches will enhance our understanding of GI variability 
across populations and time scales, enabling personalized 
interventions and population‐level dietary 
recommendations. Finally, the limited application of 
Explainable AI in GI research represents a critical gap. XAI 
tools, such as SHAP, have only been applied in two studies, 
highlighting an opportunity for future work. By integrating 
XAI into GI prediction models, researchers can improve 
transparency and trust while uncovering the factors driving 
glycemic variability, such as food preparation methods, 
ripeness, and individual metabolic responses. XAI has the 
potential to bridge the gap between advanced AI techniques 
and their practical, interpretable application in healthcare, 
paving the way for more effective and user‐centric glycemic 
management solutions. By addressing these gaps and 
leveraging advanced AI methodologies, future research can 
unlock transformative potential in Glycemic Index prediction, 
management, and personalized healthcare. 

The future of Glycemic Index-related studies can lead to 
providing personalized nutrition advice by integrating CGM 
devices, wearable sensors, and mobile health applications 
that allow AI models to learn users' real-time responses. 
These models can predict individual GI values using each 
person’s physiology, activity level, stress, and circadian 
rhythm. These AI systems can help individuals maintain a 
stable glucose level by giving personalized meal plans and 
precision dietary interventions. Another emerging avenue is 
multimodal data fusion, where different data sources such as 
food composition, metabolic responses, environmental 
context, and even emotional or behavioral cues can be used to 
build a holistic model for glycemic dynamics. The fusion can 
include different data types such as image-based meal 
recognition, nutrient text analysis, and CGM data combined to 
build context-aware predictive models to monitor glucose 
level fluctuation. 

Figure 10 indicates the main objective of the research 
covered by the GI-related studies. The significant number of 
studies dedicated to predicting Glycemic Index using machine 
learning and artificial intelligence underscores a prevailing 
trend in current research. This focus reflects the scientific 
community’s commitment to leveraging advanced 
computational methods to forecast GI values accurately, 
thereby enhancing dietary recommendations and metabolic 
health management. In contrast, the relatively limited 
research on developing low‐GI foods, with only one study [26] 
identified highlights a notable gap in the literature. 
Addressing this disparity presents a valuable opportunity for 
future investigations to concentrate on creating and 
promoting low‐GI food options. Such efforts could 
significantly contribute to dietary interventions aimed at 
improving glycemic control and reducing the risk of 
metabolic disorders. Pilot studies have demonstrated the 
feasibility of implementing low‐GI diets in primary care 
settings, suggesting that with appropriate support and 

resources, patients can successfully adopt these dietary 
changes. However, these studies also indicate the need for 
larger‐scale research to confirm the benefits and practicality 
of such interventions across diverse populations. In 
summary, while substantial progress has been made in 
predicting GI through ML and AI, there is a pressing need for 
future research to prioritize the development of low‐GI foods 
and to conduct comprehensive studies evaluating the 
effectiveness of personalized nutrition strategies. Such 
endeavors will be crucial in advancing dietary 
recommendations and improving health outcomes related to 
glycemic control. 

 
Predicting GI [13],[17],[18],[19],[20],[21],[22], 
   [23],[24],[25],[26],[27],[28],[29], 
   [30],[31],[32],[33],[34],[35],[36], 
   [37],[39],[40],[41],[60],[61],[62], 
   [63],[64],[65],[66],[67],[68],[69], 
   [70],[71],[72],[73],[74],[76],[77], 
   [78],[79],[80],[81],[82],[83],[84], 
   [85],[86],[87],[88],[89],[90],[91], 
   [92],[93],[94],[95],[96],[97],[98],  
   [99],[100],[101],[102], [115] 
 
Calculating GI  [13],[14],[16],[22],[28],[36],[37], 
    [81],[103],[109] 
 
Recent Insights  [63],[94] 
 
Analyzing GI with  [15],[24],[34]  
Food Processing 

 
Developing Low [26]  
GI Foods 
 
GL Estimation &  [25],[29],[31],[35],[38],[39],[71],  
Food   [74],[76],[82],[84],[85],[95],[97], 
Recommendation [105] 
 
Evaluating &  [38],[39],[72],[74],[76] 
Testing GL and  
GI Response 

 
Figure 10. Area covered by the research 

7. Conclusion 

The Glycemic Index serves as a vital indicator for 
understanding how foods influence blood glucose levels, 
playing a key role in managing diabetes and promoting 
healthy dietary habits. While predictive modeling dedicated 
to GI remains sparse, recent advances in machine learning 
approaches have enabled more accurate estimations of GI and 
inter‐individual glycemic responses than traditional methods. 
Deep learning techniques, in particular, have demonstrated 
their effectiveness in uncovering complex patterns within 
data, offering scalable and precise GI predictions for a wide 
range of foods. This advancement paves the way for 
significant contributions to personalized nutrition and 
dietary recommendations. Interestingly, reinforcement 
learning has not yet been extensively explored in GI‐focused 
research, marking an open area for future investigation. The 
potential of RL in this domain offers exciting opportunities to 
expand the scope of AI‐driven solutions for dietary 
management. Additionally, although image processing 
techniques were not a primary focus in this study, they have 
emerged as crucial facilitators for enhancing the application 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

122 

 

of other AI methodologies in GI research. Only two studies in 
the reviewed literature directly employed explainable AI 
techniques, specifically SHAP, to enhance the interpretability 
of their models. While other studies did not explicitly 
integrate XAI methodologies, they made efforts to validate 
and clarify how AI algorithms arrived at specific decisions or 
predictions. To achieve this, several methods were employed 
to improve model interpretability. Linear regression models, 
with their straightforward representation of the relationship 
between input features and target variables, provided clear 
insights into feature influence. Decision trees, by creating a 
series of binary choices, offered an intuitive, tree‐like 
structure to trace decision‐making paths. Similarly, rule‐
based systems used “if‐then” rules to form logical and 
transparent reasoning processes, making them valuable tools 
for understanding machine learning model outputs. Most 
studies reviewed utilize continuous glucose monitoring data 
as inputs for their models, reflecting a trend toward 
leveraging real‐time and highly granular data. Early AI‐based 
GI research predominantly relied on statistical learning 
techniques; however, with the rise of ML and advanced DL 
approaches, the field has shifted toward leveraging these 
powerful tools for deeper insights and improved accuracy. 
The majority of studies employ self‐developed datasets, as 
these datasets are specifically designed and curated to meet 
the unique objectives and requirements of the research. By 
tailoring the data to the problem at hand, researchers can 
ensure that the models are trained on highly relevant and 
domain‐specific information, which significantly improves 
their accuracy, reliability, and overall performance. This 
customized approach also allows for better control over the 
quality and diversity of the data, addressing potential gaps or 
biases that may be present in publicly available datasets. 
Consequently, the use of self‐developed datasets not only 
enhances the precision of the models but also ensures that the 
outcomes are better aligned with the intended purpose of the 
study. In conclusion, the intersection of AI and GI research is 
at an exciting juncture, with deep learning and emerging 
techniques like reinforcement learning presenting untapped 
potential. These advancements hold promise for addressing 
existing gaps in GI prediction, improving scalability, and 
enabling more personalized approaches to nutrition and 
healthcare. This review distinguishes itself by connecting 
glycemic index research with a wide range of AI techniques, 
emphasizing methodological diversity and revealing 
underexplored opportunities in explainable and 
reinforcement learning for food analytics. 

Ethical issue 
The authors are aware of and comply with best practices in 
publication ethics, specifically regarding authorship 
(avoidance of guest authorship), dual submission, 
manipulation of figures, competing interests, and compliance 
with research ethics policies. The authors adhere to 
publication requirements that the submitted work is original 
and has not been published elsewhere. 

Data availability statement 
The manuscript contains all the data. However, more data will 

be available upon request from the authors. 

Conflict of interest 

The authors declare no potential conflict of interest. 

 

 

 

References 

[1] B. R. Group, “AI and The Future of Healthcare: A 2024 
Survey,” Berkeley Research Group, 2024. [Online]. 
Available: https://media.thinkbrg.com/wp-
content/uploads/2024/02/26113735/BRG-Report-
AI-and-The-Future-of-Healthcare.pdf 

[2] World Economic Forum, “World Economic Forum; 
Annual Report 2023-2024,” World Economic Forum, 
Geneva, Switzerland, 2024. [Online]. Available: 
https://www.weforum.org/publications/annual-
report-2023-2024/ 

[3] L. Jiaying, “Cutting-edge tech boosting diagnostic 
accuracy,” China Dly., 2024, [Online]. Available: 
https://www.chinadaily.com.cn/a/202407/24/WS6
6a06492a31095c51c50fa09.html 

[4] R. Awasthi et al., “Artificial Intelligence in Healthcare: 
2023 Year in Review,” medRxiv, pp. 2024–02, 2024. 

[5] World Health Organization, “Noncommunicable 
diseases.” Sep. 13, 2024. [Online]. Available: 
https://www.who.int/news-room/fact-
sheets/detail/noncommunicable-diseases 

[6] A. Kumar, R. Gangwar, A. Ahmad Zargar, R. Kumar, and 
A. Sharma, “Prevalence of diabetes in India: A review 
of IDF diabetes atlas 10th edition,” Curr. Diabetes Rev., 
vol. 20, no. 1, pp. 105–114, 2024. 

[7] K. O. Olimjonovna, “UNDERSTANDING THE CAUSES 
AND RISK FACTORS OF DIABETES,” Biol. VA Kim. 
FANLARI ILMIY JURNALI, vol. 2, no. 5, pp. 1–7, 2024. 

[8] N. A. ElSayed et al., “Introduction and methodology: 
standards of care in diabetes—2023,” Diabetes care, 
vol. 46, no. Supplement_1. Am Diabetes Assoc, pp. S1–
S4, 2023. 

[9] N. G. Forouhi and N. J. Wareham, “Epidemiology of 
diabetes,” Medicine (Baltimore), vol. 47, no. 1, pp. 22–
27, Jan. 2019, doi: 10.1016/j.mpmed.2018.10.004. 

[10] D. Moher, A. Liberati, J. Tetzlaff, D. G. Altman, and  t 
PRISMA Group*, “Preferred reporting items for 
systematic reviews and meta-analyses: the PRISMA 
statement,” Ann. Intern. Med., vol. 151, no. 4, pp. 264–
269, 2009. 

[11] M. Mansouri, S. Benabdellah Chaouni, S. Jai Andaloussi, 
and O. Ouchetto, “MFOOD-70: Moroccan Food Dataset 
for Food Image Recognition Towards Glycemic Index 
Estimation,” in The Proceedings of the International 
Conference on Smart City Applications, Springer, 
2023, pp. 372–380. 

[12] S. Murillo et al., “Culinary strategies to manage 
glycemic response in people with type 2 diabetes: A 
narrative review,” Front. Nutr., vol. 9, p. 1025993, 
2022. 

[13] D. J. Jenkins et al., “Glycemic index of foods: a 
physiological basis for carbohydrate exchange,” Am. J. 
Clin. Nutr., vol. 34, no. 3, pp. 362–366, 1981. 

[14] G. Riccardi, G. Clemente, and R. Giacco, “Glycemic index 
of local foods and diets: the Mediterranean 
experience,” Nutr. Rev., vol. 61, no. suppl_5, pp. S56–
S60, 2003. 

[15] V. Capriles, K. Coelho, A. Guerra-Matias, and J. Arêas, 
“Effects of processing methods on amaranth starch 
digestibility and predicted glycemic index,” J. Food Sci., 
vol. 73, no. 7, pp. H160–H164, 2008. 

[16] T. M. Wolever et al., “Measuring the glycemic index of 
foods: interlaboratory study,” Am. J. Clin. Nutr., vol. 87, 
no. 1, pp. 247S-257S, 2008. 

[17] P. Hernandez, V. Rojas, and C. Mata, “Methodology for 
adding glycemic index values to a Venezuelan food 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

123 

 

composition database,” Meas. Food, vol. 7, p. 100048, 
2022. 

[18] A. Ali, M. I. Waly, M. Al-Mahrazi, J. Al-Maskari, and S. 
AlWaheibi, “Glycemic Index and Glycemic Load of 
Selected Omani Rice Dishes,” Food Sci. Eng., pp. 1–8, 
2023. 

[19] C. Monnard et al., “Nutritional Composition of Infant 
Cereal Prototypes Can Precisely Predict Their 
Glycemic Index,” Nutrients, vol. 14, no. 18, p. 3702, 
2022. 

[20] S.-H. Kim, D.-Y. Kim, S.-W. Chun, J. Kim, and J. Woo, 
“Impartial feature selection using multi-agent 
reinforcement learning for adverse glycemic event 
prediction,” Comput. Biol. Med., vol. 173, p. 108257, 
2024. 

[21] T. Miyakoshi and Y. M. Ito, “Association of Blood 
Glucose Data with Physiological and Nutritional Data 
from Dietary Surveys Investigated Using Publicly 
Available Wearable-type Databases,” JMIR Diabetes, 
Oct. 2024, doi: 10.2196/62831. 

[22] M. A. Kader et al., “Dataset on developing low glycemic 
index rice variety suitable for irrigated ecosystem in 
Bangladesh,” Data Brief, vol. 55, p. 110757, 2024. 

[23] I. Pasqualoni, R. Tolve, B. Simonato, and F. Bianchi, 
“The Impact of Selected Ingredients on the Predicted 
Glycemic Index and Technological Properties of 
Bread,” Foods, vol. 13, no. 16, p. 2488, 2024. 

[24] A. Kumar et al., “Glycemic index of Rice: Role in 
Diabetics,” NRRI Res. Bull., no. 44, 2024. 

[25] S. Gopinath, U. S, S. N. Jathin P, P. V. J, and S. 
Palaniswamy, “Glycemic Index Based Food 
Recommendation System Using Deep Learning,” in 
2024 1st International Conference on 
Communications and Computer Science (InCCCS), 
Bangalore, India: IEEE, May 2024, pp. 1–6. doi: 
10.1109/InCCCS60947.2024.10593438. 

[26] B. D. Maibam, S. Chakraborty, C. Nickhil, and S. C. Deka, 
“Effect of Euryale ferox seed shell extract addition on 
the in vitro starch digestibility and predicted glycemic 
index of wheat-based bread,” Int. J. Biol. Macromol., 
vol. 226, pp. 1066–1078, Jan. 2023, doi: 
10.1016/j.ijbiomac.2022.11.223. 

[27] T. V. Ngo, K. Kunyanee, and N. Luangsakul, “Insights 
into recent updates on factors and technologies that 
modulate the glycemic index of rice and its products,” 
Foods, vol. 12, no. 19, p. 3659, 2023. 

[28] M. H. Pipil, M. M. Hasan, A. Setu, S. R. Ray, S. Sharmin, 
and M. J. H. Bhuiyan, “Evaluation of Glycemic Index of 
Selected Black Rice Cultivars in Bangladesh: 
Evaluation of glycemic index of selected local rice 
cultivars in Bangladesh,” J. Bangladesh Agric. Univ., 
vol. 22, no. 2, pp. 249–256, 2024. 

[29] G. Annuzzi et al., “Impact of nutritional factors in blood 
glucose prediction in type 1 diabetes through machine 
learning,” IEEE Access, vol. 11, pp. 17104–17115, 
2023. 

[30] S. Shahakar, P. Chopde, N. Purohit, A. Vishwakarma, A. 
Nite, and A. K. Sharma, “A Novel Machine Learning and 
Deep Learning Driven Prediction for Pre-diabetic 
Patients,” in 2023 6th International Conference on 
Information Systems and Computer Networks 
(ISCON), IEEE, 2023, pp. 1–6. 

[31] Z.-F. Chen, J. D. Kusuma, and S.-Y. P. K. Shiao, 
“Validating Healthy Eating Index, Glycemic Index, and 
Glycemic Load with Modern Diets for E-Health Era,” 
Nutrients, vol. 15, no. 5, p. 1263, 2023. 

[32] A. T. Sipe, M. L. Neuhouser, K. L. Breymeyer, and K. M. 
Utzschneider, “Effect of dietary glycemic index on β-
cell function in prediabetes: A randomized controlled 
feeding study,” Nutrients, vol. 14, no. 4, p. 887, 2022. 

[33] A. Rytz et al., “Predicting glycemic index and glycemic 
load from macronutrients to accelerate development 
of foods and beverages with lower glucose responses,” 
Nutrients, vol. 11, no. 5, p. 1172, 2019. 

[34] P. H. S. de Rossi, S. M. Barbalho, M. Oshiiwa, and A. A. J. 
Sampaio, “Glycemic index of foods: A review,” 
Pubsaúde, vol. 4, p. a072, 2020, doi: 
10.31533/pubsaude4.a072. 

[35] M. A. Samir, Z. A. Mohamed, M. A. A. Hussein, and A. 
Atia, “Applying Deep Learning to Track Food 
Consumption and Human Activity for Non-intrusive 
Blood Glucose Monitoring,” in 2021 IEEE 12th Annual 
Ubiquitous Computing, Electronics & Mobile 
Communication Conference (UEMCON), IEEE, 2021, 
pp. 0319–0324. 

[36] U. Gamage, D. Kumari, R. Senavirathna, and D. Rathish, 
“In vitro Starch Hydrolysis Indices of selected Sri 
Lankan Starchy Tubers.,” J. Agric. Sci. Sri Lanka, vol. 16, 
no. 2, 2021. 

[37] F. S. Atkinson, J. C. Brand-Miller, K. Foster-Powell, A. E. 
Buyken, and J. Goletzke, “International tables of 
glycemic index and glycemic load values 2021: a 
systematic review,” Am. J. Clin. Nutr., vol. 114, no. 5, pp. 
1625–1632, 2021. 

[38] W. Wongniyomkaset, N. Rungraung, N. Muangpracha, 
T. Winuprasith, and D. Trachootham, “Baseline Insulin 
Levels may Predict Response to Low Glycemic Index 
Complete Nutrition Formula: A Randomized Cross-
Over Control Trial,” 2021. 

[39] E. Pustozerov et al., “The role of glycemic index and 
glycemic load in the development of real-time 
postprandial glycemic response prediction models for 
patients with gestational diabetes,” Nutrients, vol. 12, 
no. 2, p. 302, 2020. 

[40] S. Pirasath, S. Balakumar, and V. Arasaratnam, 
“Glycemic index of traditional foods in Northern Sri 
Lanka,” 2015. 

[41] L. S. Augustin et al., “Glycemic index, glycemic load and 
glycemic response: an International Scientific 
Consensus Summit from the International 
Carbohydrate Quality Consortium (ICQC),” Nutr. 
Metab. Cardiovasc. Dis., vol. 25, no. 9, pp. 795–815, 
2015. 

[42] B. J. Venn, M. Kataoka, and J. Mann, “The use of 
different reference foods in determining the glycemic 
index of starchy and non-starchy test foods,” Nutr. J., 
vol. 13, pp. 1–6, 2014. 

[43] N. Gibson, H. C. Schönfeldt, and B. Pretorius, 
“Development of a rapid assessment method for the 
prediction of the glycemic index,” J. Food Compos. 
Anal., vol. 24, no. 4–5, pp. 750–754, 2011. 

[44] R. L. Magaletta, S. N. DiCataldo, D. Liu, H. L. Li, R. P. 
Borwankar, and M. C. Martini, “In vitro method for 
predicting glycemic index of foods using simulated 
digestion and an artificial neural network,” Cereal 
Chem., vol. 87, no. 4, pp. 363–369, 2010. 

[45] B. Mahesh, “Machine learning algorithms-a review,” 
Int. J. Sci. Res. IJSRInternet, vol. 9, no. 1, pp. 381–386, 
2020. 

[46] A. Géron, Hands-on machine learning with Scikit-
Learn, Keras, and TensorFlow.  O’Reilly Media, Inc., 
2022. 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

124 

 

[47] S. Gronauer and K. Diepold, “Multi-agent deep 
reinforcement learning: a survey,” Artif. Intell. Rev., 
vol. 55, no. 2, pp. 895–943, 2022. 

[48] H. Askr, E. Elgeldawi, H. Aboul Ella, Y. A. Elshaier, M. M. 
Gomaa, and A. E. Hassanien, “Deep learning in drug 
discovery: an integrative review and future 
challenges,” Artif. Intell. Rev., vol. 56, no. 7, pp. 5975–
6037, 2023. 

[49] K. Suzuki, “Overview of deep learning in medical 
imaging,” Radiol. Phys. Technol., vol. 10, no. 3, pp. 257–
273, 2017. 

[50] M. Bakator and D. Radosav, “Deep learning and 
medical diagnosis: A review of literature,” Multimodal 
Technol. Interact., vol. 2, no. 3, p. 47, 2018. 

[51] Q. Rao and J. Frtunikj, “Deep learning for self-driving 
cars: Chances and challenges,” in Proceedings of the 
1st international workshop on software engineering 
for AI in autonomous systems, 2018, pp. 35–38. 

[52] T. Doleck, D. J. Lemay, R. B. Basnet, and P. Bazelais, 
“Predictive analytics in education: a comparison of 
deep learning frameworks,” Educ. Inf. Technol., vol. 25, 
pp. 1951–1963, 2020. 

[53] L. Solorzano, “Image Processing, Machine Learning 
and Visualization for Tissue Analysis,” PhD Thesis, 
Acta Universitatis Upsaliensis, 2021. 

[54] S. Vajjala, B. Majumder, A. Gupta, and H. Surana, 
Practical natural language processing: a 
comprehensive guide to building real-world NLP 
systems. O’Reilly Media, 2020. 

[55] A. Saranya and R. Subhashini, “A systematic review of 
Explainable Artificial Intelligence models and 
applications: Recent developments and future trends,” 
Decis. Anal. J., vol. 7, p. 100230, 2023. 

[56] S. Lundberg, “A unified approach to interpreting model 
predictions,” ArXiv Prepr. ArXiv170507874, 2017. 

[57] “ANOVA (analysis of variance),” Statistics Solutions. 
Dec. 2010. [Online]. Available: 
https://www.statisticssolutions.com/free-
resources/directory-of-statistical-analyses/anova/ 

[58] “Tukey HSD (Honestly Significant Difference).” 
[Online]. Available: https://real-statistics.com/one-
way-analysis-of-variance-anova/unplanned-
comparisons/tukey-hsd/ 

[59] XM for Employee Experience, “T-test theory for 
surveys: An introduction,” Qualtrics. Sep. 2020. 
[Online]. Available: 
https://www.qualtrics.com/experience-
management/research/t-test-analysis/ 

[60] D. Zeevi et al., “Personalized nutrition by prediction of 
glycemic responses,” Cell, vol. 163, no. 5, pp. 1079–
1094, 2015. 

[61] M.-H. Lai et al., “Predicted glycemic index and glycemic 
index of rice varieties grown in Taiwan,” Cereal Chem., 
vol. 93, no. 2, pp. 150–155, 2016. 

[62] S. Sampath, P. Tkachenko, E. Renard, and S. V. 
Pereverzev, “Glycemic control indices and their 
aggregation in the prediction of nocturnal 
hypoglycemia from intermittent blood glucose 
measurements,” J. Diabetes Sci. Technol., vol. 10, no. 6, 
pp. 1245–1250, 2016. 

[63] J. Li and O. Arandjelovic, “Glycaemic index prediction: 
a pilot study of data linkage challenges and the 
application of machine learning,” in 2017 IEEE EMBS 
International Conference on Biomedical & Health 
Informatics (BHI), IEEE, 2017, pp. 357–360. 

[64] E. N. Okafor et al., “Cassava flour substitution 
modulates glycemic responses and glycemic index of 
wheat breads in apparent healthy volunteers,” J. Diet. 
Suppl., vol. 14, no. 4, pp. 446–452, 2017. 

[65] A. Fardet, S. Lakhssassi, and A. Briffaz, “Beyond 
nutrient-based food indices: a data mining approach to 
search for a quantitative holistic index reflecting the 
degree of food processing and including 
physicochemical properties,” Food Funct., vol. 9, no. 1, 
pp. 561–572, 2018. 

[66] S. P. K. Shiao, J. Grayson, A. Lie, and C. H. Yu, “Predictors 
of the healthy eating index and glycemic index in 
multi-ethnic colorectal cancer families,” Nutrients, vol. 
10, no. 6, p. 674, 2018. 

[67] S. K. Kumari and J. Mathana, “Blood sugar level 
indication through chewing and swallowing from 
acoustic MEMS sensor and deep learning algorithm for 
diabetic management,” J. Med. Syst., vol. 43, pp. 1–9, 
2019. 

[68] M. Ghazal, M. Alhalabi, M. Chabnari, L. Ghazal, and L. 
Fraiwan, “Design and implementation of a mobile 
device for blood glucose level assessment,” in 2018 6th 
International Conference on Future Internet of Things 
and Cloud Workshops (FiCloudW), IEEE, 2018, pp. 
134–139. 

[69] J. J. R. Barata, R. Munoz, R. D. D. C. Silva, J. J. Rodrigues, 
and V. H. C. De Albuquerque, “Internet of Things based 
on electronic and mobile health systems for blood 
glucose continuous monitoring and management,” 
IEEE Access, vol. 7, pp. 175116–175125, 2019. 

[70] S. Buranapin, “A STUDY OF THE GLYCEMIC INDEX OF 
THE MEDICAL FOOD NEO-MUNE,” J. Southeast Asian 
Med. Res., vol. 2, no. 2, pp. 67–75, 2018. 

[71] W. Seo, Y.-B. Lee, S. Lee, S.-M. Jin, and S.-M. Park, “A 
machine-learning approach to predict postprandial 
hypoglycemia,” BMC Med. Inform. Decis. Mak., vol. 19, 
pp. 1–13, 2019. 

[72] H. H. Lee, P. Achananuparp, Y. Liu, E.-P. Lim, and L. R. 
Varshney, “Estimating glycemic impact of cooking 
recipes via online crowdsourcing and machine 
learning,” in Proceedings of the 9th International 
Conference on Digital Public Health, 2019, pp. 31–35. 

[73] M. Mayo, L. Chepulis, and R. G. Paul, “Glycemic-aware 
metrics and oversampling techniques for predicting 
blood glucose levels using machine learning,” PloS 
One, vol. 14, no. 12, p. e0225613, 2019. 

[74] I. K. Mohan et al., “Application of multiple linear 
regression and machine learning algorithms to 
elucidate the association of poor glycemic control and 
hyperhomocysteinemia with microalbuminuria,” 
Indian J. Biochem. Biophys. IJBB, vol. 56, no. 2, pp. 150–
154, 2019. 

[75] I. Rodríguez-Rodríguez, J.-V. Rodríguez, I. 
Chatzigiannakis, and M. A. Zamora Izquierdo, “On the 
possibility of predicting glycaemia ‘on the fly’with 
constrained IoT devices in type 1 diabetes mellitus 
patients,” Sensors, vol. 19, no. 20, p. 4538, 2019. 

[76] J. C. Peiró, “Comparing artificial pancreas controlled by 
hybrid" closed-loop" machine learning (ML) trained 
algorithm to multi-daily injection (MDI), insulin pump 
without CGM and" sensor assisted" insulin pump 
therapies for Diabetes Type 1 (DT1) treatment,” in 
2020 International Conference on Data Analytics for 
Business and Industry: Way Towards a Sustainable 
Economy (ICDABI), IEEE, 2020, pp. 1–6. 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

125 

 

[77] S. Rizwana and M. K. Hazarika, “Application of near-
infrared spectroscopy for rice characterization using 
machine learning,” J. Inst. Eng. India Ser. A, vol. 101, no. 
4, pp. 579–587, 2020. 

[78] E. A. Pustozerov et al., “Machine learning approach for 
postprandial blood glucose prediction in gestational 
diabetes mellitus,” Ieee Access, vol. 8, pp. 219308–
219321, 2020. 

[79] H. Hesse, R. Frozza, V. Corbellini, C. Reuter, and M. 
Burgos, “The use of Artificial Neural Network for lipid 
and glycaemic profiles quantification through infrared 
spectroscopy,” in Anais do XVI Workshop de 
Informática Médica, SBC, 2016, pp. 2613–2616. 

[80] B. Olawoye, S. O. Gbadamosi, I. O. Otemuyiwa, and C. T. 
Akanbi, “Gluten-free cookies with low glycemic index 
and glycemic load: optimization of the process 
variables via response surface methodology and 
artificial neural network,” Heliyon, vol. 6, no. 10, 2020. 

[81] N. T. D. Anh et al., “Determining the Glycemic Index of 
Nutritional Product for Diabetes Mellitus-Np through 
Measuring Glycemic Responses to Reference Food 
(Glucose) and Test Food (Nutritional product-Np),” J. 
Pharm. Res. Int., vol. 33, no. 47B, pp. 298–307, 2021. 

[82] A. Das et al., “A sparse coding approach to automatic 
diet monitoring with continuous glucose monitors,” in 
ICASSP 2021-2021 IEEE International Conference on 
Acoustics, Speech and Signal Processing (ICASSP), 
IEEE, 2021, pp. 2900–2904. 

[83] G. Lazzarinetti, N. Massarenti, and S. Mantisi, “A Cloud-
Hosted Online Learning Approach for Glycemic Index 
Forecasting.,” in AIABI@ AI* IA, 2021. 

[84] S. Narmadha, S. Gokulan, M. Pavithra, R. Rajmohan, and 
T. Ananthkumar, “Determination of various deep 
learning parameters to predict heart disease for 
diabetes patients,” in 2020 International Conference 
on System, Computation, Automation and Networking 
(ICSCAN), IEEE, 2020, pp. 1–6. 

[85] E. M. Kariuki, “Food recommender system for Diabetes 
Type 2 patients,” PhD Thesis, Strathmore University, 
2021. 

[86] L. Martínez-Delgado, M. Munoz-Organero, and P. 
Queipo-Alvarez, “Using absorption models for insulin 
and carbohydrates and deep leaning to improve 
glucose level predictions,” Sensors, vol. 21, no. 16, p. 
5273, 2021. 

[87] B. Bent et al., “Non-invasive wearables for remote 
monitoring of HbA1c and glucose variability: proof of 
concept,” BMJ Open Diabetes Res. Care, vol. 9, no. 1, p. 
e002027, 2021. 

[88] M. S. Patel et al., “Predicting changes in glycemic 
control among adults with prediabetes from activity 
patterns collected by wearable devices,” Npj Digit. 
Med., vol. 4, no. 1, p. 172, 2021. 

[89] A. Massaro, N. Magaletti, G. Cosoli, A. Leogrande, and F. 
Cannone, “Use of machine learning to predict the 
glycemic status of patients with diabetes,” in Medical 
Sciences Forum, MDPI, 2022, p. 11. 

[90] B. Neamtu, M. O. Negrea, and I. Neagu, “Predicting 
Glycemic Control in a Small Cohort of Children with 
Type 1 Diabetes Using Machine Learning Algorithms,” 
Mathematics, vol. 11, no. 20, p. 4388, 2023. 

[91] T. M. Abegaz, M. Ahmed, F. Sherbeny, V. Diaby, H. Chi, 
and A. A. Ali, “Application of machine learning 
algorithms to predict uncontrolled diabetes using the 
All of Us research program data,” in Healthcare, MDPI, 
2023, p. 1138. 

[92] S. Romero-Tapiador et al., “AI4FoodDB: a database for 
personalized e-Health nutrition and lifestyle through 
wearable devices and artificial intelligence,” Database, 
vol. 2023, p. baad049, 2023. 

[93] Y. Liu, Z. Han, X. Liu, J. Wang, C. Wang, and R. Liu, 
“Estimation method and research of fruit glycemic 
load index based on the fusion SE module Faster R-
CNN,” Comput. Electr. Eng., vol. 109, p. 108696, 2023. 

[94] P. G. Jacobs et al., “Artificial intelligence and machine 
learning for improving glycemic control in diabetes: 
best practices, pitfalls and opportunities,” IEEE Rev. 
Biomed. Eng., 2023. 

[95] A. A. Crystal, M. Valero, V. Nino, and K. H. Ingram, 
“Empowering Diabetics: Advancements in 
Smartphone-Based Food Classification, Volume 
Measurement, and Nutritional Estimation,” Sensors, 
vol. 24, no. 13, p. 4089, 2024. 

[96] D. Sisodia and D. S. Sisodia, “Prediction of diabetes 
using classification algorithms,” Procedia Comput. Sci., 
vol. 132, pp. 1578–1585, 2018. 

[97] M. Ibrahim, A. Beneyto, I. Contreras, and J. Vehi, “An 
ensemble machine learning approach for the detection 
of unannounced meals to enhance postprandial 
glucose control,” Comput. Biol. Med., vol. 171, p. 
108154, 2024. 

[98] K. A. Della Corte, D. Della Corte, S. Titensor, B. Yang, 
and S. Liu, “Development of a national database for 
dietary glycemic index and load for nutritional 
epidemiologic studies in the United States,” Am. J. Clin. 
Nutr., 2024. 

[99] X. Wang, W. Wang, H. Ren, X. Li, and Y. Wen, “Prediction 
and analysis of risk factors for diabetic retinopathy 
based on machine learning and interpretable models,” 
Heliyon, vol. 10, no. 9, 2024. 

[100] R. Bouchi et al., “Internet of things-based approach for 
glycemic control in people with type 2 diabetes: A 
randomized controlled trial,” J. Diabetes Investig., 
2024. 

[101] M. Mansour, M. S. Darweesh, and A. Soltan, “Wearable 
devices for glucose monitoring: A review of state-of-
the-art technologies and emerging trends,” Alex. Eng. 
J., vol. 89, pp. 224–243, 2024. 

[102] Y. Zhang, X. Bao, Y. Zhu, Z. Dai, Q. Shen, and Y. Xue, 
“Advances in Machine Learning Screening of Food 
Bioactive Compounds,” Trends Food Sci. Technol., p. 
104578, 2024. 

[103] R. Chlup et al., “Automated computation of glycemic 
index for foodstuffs using continuous glucose 
monitoring,” J. Diabetes Sci. Technol., vol. 2, no. 1, pp. 
67–75, 2008. 

[104] S. Al Hamli, M. Holmes, S. Khokhar, and C. Orfila, 
“Predicting glyceamic index in cereal and legume-
based foods from macronutrient composition data,” 
Proc. Nutr. Soc., vol. 70, no. OCE4, p. E122, 2011. 

[105] A. J. Jara, M. A. Zamora, and A. F. Skarmeta, “An internet 
of things–based personal device for diabetes therapy 
management in ambient assisted living (AAL),” Pers. 
Ubiquitous Comput., vol. 15, pp. 431–440, 2011. 

[106] E. Bas, “A robust optimization approach to diet 
problem with overall glycemic load as objective 
function,” Appl. Math. Model., vol. 38, no. 19–20, pp. 
4926–4940, 2014. 

[107] N. Struble, “Measuring Glycemic Variability and 
Predicting Blood Glucose Levels Using Machine 
Learning Regression Models,” Master’s Thesis, Ohio 
University, 2013. 



NH. Wanigasingha et al. /Future Technology                                                                 February 2026| Volume 05 | Issue 01 | Pages 93-126 

126 

 

[108] A. Flint et al., “The use of glycaemic index tables to 
predict glycaemic index of composite breakfast 
meals,” Br. J. Nutr., vol. 91, no. 6, pp. 979–989, 2004. 

[109] T. M. Wolever, D. J. Jenkins, V. Vuksan, R. G. Josse, G. S. 
Wong, and A. L. Jenkins, “Glycemic index of foods in 
individual subjects,” Diabetes Care, vol. 13, no. 2, pp. 
126–132, 1990. 

[110] GI Group and A. Barclay, “Glycemic index – glycemic 
index research and GI news.” Nov. 2024. [Online]. 
Available: http://www.glycemicindex.com 

[111] P. Rani, R. Lamba, R. K. Sachdeva, P. Bathla, and A. N. 
Aledaily, “Diabetes prediction using machine learning 
classification algorithms,” in 2023 International 
Conference on Smart Computing and Application 
(ICSCA), IEEE, 2023, pp. 1–5. 

[112] G. Annuzzi et al., “Exploring Nutritional Influence on 
Blood Glucose Forecasting for Type 1 Diabetes Using 
Explainable AI,” IEEE J. Biomed. Health Inform., vol. 28, 
no. 5, pp. 3123–3133, May 2024, doi: 
10.1109/JBHI.2023.3348334. 

[113] D. L. Trout, K. M. Behall, and O. Osilesi, “Prediction of 
glycemic index for starchy foods,” Am. J. Clin. Nutr., vol. 
58, no. 6, pp. 873–878, 1993. 

[114] T. Larsen et al., “The Diet, Obesity and Genes 
(Diogenes) Dietary Study in eight European countries–
a comprehensive design for long-term intervention,” 
Obes. Rev., vol. 11, no. 1, pp. 76–91, 2010. 

[115] I. Rodríguez-Rodríguez, J.-V. Rodríguez, I. 
Chatzigiannakis, and M. A. Zamora Izquierdo, “On the 
possibility of predicting glycaemia ‘on the fly’with 
constrained IoT devices in type 1 diabetes mellitus 
patients,” Sensors, vol. 19, no. 20, p. 4538, 2019. 

 
 

  

This article is an open-access article distributed under the 

terms and conditions of the Creative Commons Attribution 

(CC BY) license 

(https://creativecommons.org/licenses/by/4.0/). 

https://creativecommons.org/licenses/by/4.0/

