










































CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

5 
      
 

Abstract— In the use of IoMT-enabled diagnostic tools, this research seeks to improve the accuracy of skin disease identification. For 

texture analysis, the K-Nearest Neighbours (KNN) method is used, and Class Activation Mapping (CAM) is used for visual interpretability. 

Image quality is enhanced by preprocessing techniques like DF-U-Net segmentation, and data imbalance is addressed with SMOTE. 

Combining the explainability of CAM with the efficiency of KNN, the system achieves 99.18% accuracy, outperforming current techniques 

and offering dependable diagnostic support for medical practitioners. 

Objective: Using KNN for texture-based classification, optimising the K-value, and adding CAM for improved model interpretability, we will 

create an IoMT-enabled system for precise skin disease diagnosis while resolving data imbalance for better diagnostic performance. 

Methods: IoMT-enabled image acquisition serves as the system's foundation. DF-U-Net segmentation, contrast enhancement, and noise 

reduction are examples of preprocessing. Texture features are extracted using the Gray-Level Co-occurrence Matrix (GLCM), KNN is cross-

validation optimised, and the dataset is balanced using SMOTE. Important picture regions for classification are highlighted by CAM. 

Results: The system outperformed XGBoost and EFF_D_SVM with an accuracy of 99.18% and a false-negative rate (FNR) of 0.48%. 

Conclusion: Combining KNN and CAM results in a skin disease detection system that is reliable and easy to understand, with good 

performance for real-world medical applications. 

Keywords: IoMT, K-Nearest Neighbors, Class Activation Mapping, Skin Disease Detection, DF-U-Net, SMOTE, Medical Image Processing. 

 

Chinese Traditional Medical Journal 
 

 

Accurate Skin Disease Detection with K-Nearest Neighbors and 

CAM in IoMT-Enabled Diagnostic Solutions 

Surendar Rama Sitaraman,Intel Corporation, California, USA  

surendar.rama.sitaraman@ieee.org 

Poovendran Alagarsundaram, Humetis Technologies Inc, Kingston, NJ, USA 

poovendrana@ieee.org 

Veerandra Kumar R, Department of Computer Science and Engineering, Panimalar Engineering 

College, Chennai, India. veerandrakumar.r@panimalar.ac.in 

Dr. Dede Kurniadi Associate Professor of Computer Science Department of Computer Science 

Institut Teknologi Garut Jl. Mayor Syamsu No. 1, Garut, Indonesia 

dede.kurniadi@itg.ac.id 

 

1. INTRODUCTION 

A person's lifestyle, environment, parasites, 

viruses, bacteria, fungi, infections, and 

infections can all contribute to skin 

problems. If untreated, these illnesses can 

have a negative impact on a person's health. 

For prompt action and appropriate treatment 

to be guaranteed, early and precise detection 

is crucial. The development of Internet of 

mailto:surendar.rama.sitaraman@ieee.org
mailto:poovendrana@ieee.org
mailto:veerandrakumar.r@panimalar.ac.in


CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

6 
      
 

Medical Things (IoMT) technology presents 

a viable path towards the creation of more 

precise diagnostic tools. This work attempts 

to improve the accuracy of skin disease 

identification by utilising machine learning 

techniques such as the K-Nearest 

Neighbours (KNN) algorithm. Image texture 

features are used to analyse and categorise 

skin conditions using KNN, which is well-

known for its simplicity and efficacy, in 

conjunction with Class Activation Mapping 

(CAM). The system seeks to resolve issues 

like data imbalance and determine the ideal 

"K" value in order to deliver more 

dependable diagnoses for practical 

healthcare applications. 

In recent years, there has been a substantial 

evolution in the field of skin disease 

identification. Dermatologists use traditional 

diagnostic techniques, which mostly rely on 

visual assessment. When necessary, clinical 

tests are then performed. However, 

computer-aided diagnostic (CAD) systems 

have become more effective instruments for 

early detection and diagnosis as a result of 

advancements in medical technology. 

Especially for more serious disorders like 

melanoma, these systems use machine 

learning and image processing approaches to 

increase diagnostic efficiency and accuracy 

(Oumarou, 2023; Esti et al., 2023). 

The accuracy of diagnosing skin diseases 

has greatly increased recently because to 

developments in machine learning 

algorithms and IoMT-enabled diagnostic 

tools. When paired with machine learning 

methods like K-Nearest Neighbours (KNN), 

the massive volumes of real-time medical 

data collected by IoMT devices enable more 

accurate classification of skin conditions 

(Pandey et al., 2023). Class Activation 

Mapping (CAM) integration improves the 

explainability of the model by offering a 

visual representation of the regions of 

interest in medical imaging. These 

technological advancements represent a 

quantum leap in healthcare, offering faster, 

more accessible, more accurate diagnostic 

tools to support medical practitioners and 

lower the number of misdiagnoses (Zhang et 

al., 2023). 

The following are the main goals of this 

study: 

➢ To build an IoMT-enabled diagnostic 

system for precise identification of 

skin conditions. 

➢ Putting into practice the K-Nearest 

Neighbours (KNN) algorithm for 

using texture information in images 

to classify skin disorders. 

➢ Determine the ideal "K" value for the 

KNN model in order to improve 

classification accuracy. 

➢ Using Class Activation Mapping 

(CAM) to enhance the diagnostics of 

skin diseases' interpretability. 

➢ Solve problems with data imbalance 

and enhance the generalisation and 

resilience of the model for practical 

applications (Santana-Mancilla et 

al., 2023; Muneer and Khan, 2023). 

Skin disease detection algorithms' accuracy 

and reliability are limited by issues such as 

data imbalance and performance fluctuation, 

even with advances in machine learning and 

IoMT. When dealing with under-represented 

illness classes, in particular, many models 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

7 
      
 

find it difficult to generalise across various 

datasets, which might result in biassed 

results, (Oumarou, 2023). Furthermore, the 

necessity for more reliable and consistent 

models is suggested by the heterogeneity in 

performance under various settings. Future 

studies should concentrate on resolving 

these problems by optimising 

hyperparameters and using approaches like 

ensemble learning or oversampling to 

guarantee greater stability and dependability 

in the categorisation of skin diseases (Esti et 

al., 2023). 

2. LITERATURE SURVEY 

Surendar Rama Sitaraman (2021) examines 

Crow Search Optimization (CSO), which 

aims to enhance AI-driven disease diagnosis 

in smart healthcare. To improve accuracy, 

precision, recall, and F1-score, CSO 

optimizes CNNs and LSTMs, outperforming 

conventional techniques like genetic 

algorithms and particle swarm optimization. 

It can be used in a wide range of healthcare 

applications, and additional research will 

concentrate on real-time implementation and 

ethics. 

In order to effectively classify different 

forms of brain tumours, Zhang et al. (2023) 

provide EFF_D_SVM, a robust multi-type 

brain tumour classification system that 

makes use of an upgraded support vector 

machine (SVM). In order to provide more 

accurate diagnoses and treatment plans for 

patients with brain tumours, the system 

focusses on increasing accuracy and 

computing efficiency. 

Rezaee et al. (2022) suggest a method for 

detecting falls that makes use of deep 

transfer learning and IoMT-enabled thermal 

imaging. The technology improves real-time 

surveillance accuracy through the 

integration of big heath data. Its main 

objective is to identify falls in healthcare 

more effectively by employing IoMT and 

thermal imagery to enhance decision-

making and pervasive monitoring. 

The book Data Modelling and Analytics for 

the Internet of Medical Things by Pandey et 

al. (2023) explores analytics and data 

modelling frameworks specifically designed 

for IoMT. The book focusses on large-scale 

data management and how it may be used to 

improve patient care and system efficiency. 

It also discusses how to integrate, analyse, 

and analyse medical data to improve 

healthcare delivery. 

Santana-Mancilla et al. (2023) provide a 

technique that uses supervised machine 

learning on respiratory frequency data 

obtained from IoMT devices to predict 

aberrant respiratory patterns in older 

persons. The study improves early 

respiratory issue diagnosis by utilising 

machine learning algorithms, allowing for 

prompt interventions for improved 

healthcare management in older populations. 

Surendar (2021) explores the way healthcare 

data is handled and utilized is being 

revolutionized by AI-driven healthcare 

systems that are strengthened by mobile 

computing and advanced data analytics. 

Important areas like data collecting, 

processing, storage, and application 

development are the focus of the study. 

These systems facilitate real-time analysis, 

predictive models, and customized 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

8 
      
 

healthcare services by combining 

technologies such as distributed file storage, 

NoSQL databases, and parallel computing. 

According to the research, AI greatly 

increases healthcare accuracy, speed, and 

dependability, which enhances patient care 

and operational effectiveness. 

An extensive assessment of wearable and 

mobile sensors for data-driven health 

monitoring systems is given by Anikwe et 

al. (2022), with an emphasis on recent 

developments and potential future 

applications. In addition to discussing how 

these sensors can be integrated with data 

analytics to improve real-time health 

monitoring, the study looks into the 

possibility of AI-driven personalised 

healthcare in the future. 

Using machine learning approaches, Muneer 

and Khan (2023) present an intelligent 

disease prediction system for Hepatitis C 

patients that predicts the course of the 

illness. With data-driven decision-making, 

the system improves early diagnosis, 

facilitating more efficient treatment and 

individualised healthcare management, and 

eventually leading to better patient 

outcomes. 

K-nearest neighbours (KNN) method is used 

by Araaf and Nugroho (2023) to analyse and 

diagnose skin disorders based on textural 

information in images. Their strategy is to 

enable more accurate detection of different 

skin conditions by utilising texture-based 

image analysis to improve diagnostic 

accuracy. This approach improves the 

efficiency of diagnosing skin conditions and 

leads to more favourable treatment results. 

Surendar (2020) investigates the impact of 

incorporating Big Data Analytics and 

Artificial Intelligence (AI) into mobile 

health (m-Health) technology on healthcare 

delivery. It demonstrates the remarkable 

92% accuracy of neural networks in 

handling intricate medical data and 

highlights how crucial Apache Spark and 

Hadoop are to enabling fast data processing 

for prompt medical actions. Despite these 

developments, there are still issues with 

handling unstructured data from wearable 

technologies and safeguarding data privacy, 

which emphasizes the need for more study 

and development to fully exploit the 

potential of big data and artificial 

intelligence in healthcare. 
 

3. METHODOLOGY 

The suggested IoMT-enabled diagnostic 

system for identifying skin diseases 

combines Class Activation Mapping (CAM) 

for interpretability with machine learning 

methods, specifically the K-Nearest 

Neighbours (KNN) algorithm. The 

technique describes a methodical procedure 

that starts with the collecting of input data, 

moves on to image processing, feature 

extraction, classification, and explainability. 

To improve generalisability and accuracy, 

alternative approaches are included, 

guaranteeing the system's efficacy in actual 

healthcare situations. 

3.1. Input: Skin Disease Image 

Acquisition via IoMT Devices 

Using IoMT-capable equipment like 

dermatoscopes, mobile cameras, or 

specialised imaging sensors, the proposed 

system's initial stage entails taking pictures 

of skin diseases. These gadgets have the 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

9 
      
 

capacity to capture sharp photographs in real 

time and securely send the information to a 

cloud-based server for analysis. 

Apart from picture data, the system gathers 

clinical metadata such patient 

characteristics, symptoms, and past medical 

records. The following inputs are indicated: 

𝐼 = {𝐼1, 𝐼2, … , 𝐼𝑛}                                                         

(1) 

where 𝐼𝑖 denotes each skin disease image in 

the dataset. 

3.2. Preprocessing: Image 

Standardization and Enhancement 

In order to guarantee consistency and 

enhance image quality for reliable analysis, 

the subsequent preprocessing procedures are 

carried out: 

Resizing: To ensure effective processing, all 

photos are shrunk to a uniform dimension to 

standardise the input size. The definition of 

the resizing function is: 

𝐼𝑟 = resize⁡(𝐼, 𝑤, ℎ)                                                            

(2) 

where 𝑤 and ℎ represent the new width and 

height of the images. 

Noise Reduction: Medical imaging benefits 

greatly from the application of the Bilateral 

Filter, which eliminates noise while 

maintaining image edges. 

𝐼𝑓 = bilateralFilter⁡(𝐼𝑟)                                                       

(3) 

where 𝐼𝑓 represents the filtered image. 

Contrast Enhancement: Contrast 

enhancement is performed using CLAHE 

(Contrast Limited Adaptive Histogram 

Equalization), ensuring that the critical 

features of the images are highlighted: 

𝐼𝑐𝑒 = CLAHE⁡(𝐼𝑓)                                                           

(4) 

Segmentation using DF-U-Net: The skin 

lesion segmentation is carried out using the 

DropFilter U-Net (DF-U-Net) design. By 

incorporating regularisation approaches to 

prevent overfitting and accurately define the 

lesion borders, DF-U-Net enhances 

conventional U-Net. The divided picture is 

represented by: 

𝑆 = DF − U − Net(𝐼𝑐𝑒)                                                      

(5) 

In addition to minimising the effect of 

background noise, this method successfully 

locates the region of interest, or skin lesion. 

3.3. Feature Extraction: Texture and 

Morphological Features 

The skin disease is characterised by 

extracting texture and morphological 

information from the segmented lesion. The 

system extracts important texture properties, 

such as contrast, correlation, energy, and 

homogeneity, using the Gray-Level Co-

occurrence Matrix (GLCM): 

Contrast: 

𝐶 = ∑  𝑖,𝑗 (𝑖 − 𝑗)2𝑃(𝑖, 𝑗)                                                       

(6) 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

10 
      
 

Correlation: 

Corr =
∑  𝑖,𝑗  (𝑖−𝜇𝑖)(𝑗−𝜇𝑗)𝑃(𝑖,𝑗)

𝜎𝑖𝜎𝑗
                                                   

(7) 

Energy: 

𝐸 = ∑  𝑖,𝑗 𝑃(𝑖, 𝑗)
2                                                         

(8) 

Homogeneity: 

𝐻 = ∑  𝑖,𝑗
𝑃(𝑖,𝑗)

1+|𝑖−𝑗|
                                                         

(9) 

where 𝑃(𝑖, 𝑗) is the co-occurrence matrix 

value for the pixel intensities 𝑖 and 𝑗, and 𝜇 

and 𝜎 stand for the mean and standard 

deviation of the intensity values. 

Morphological operations are utilised to 

extract lesion shape and boundary features, 

which offer supplementary indicators for 

precise diagnosis, in addition to texture 

features. 

3.4. Classification: K-Nearest Neighbors 

(KNN) Algorithm 

By utilising the retrieved morphological and 

texture information, the K-Nearest 

Neighbours (KNN) algorithm is utilised for 

classification. To compare the new input 

sample to the K-nearest neighbours in the 

training set, KNN uses the Euclidean 

distance as the similarity metric. 

𝑑(𝑋𝑡, 𝑋𝑠) = √∑  𝑛
𝑖=1   (𝑋𝑡,𝑖 − 𝑋𝑠,𝑖)

2
                                                        

(10) 

where 𝑋𝑠 is a training sample and 𝑋𝑡 is the 

test sample. Cross-validation is used to 

determine the ideal value of K in order to 

reduce bias and variation. Based on the 

majority vote of the closest neighbours, the 

KNN algorithm returns the skin disease 

class that is most likely to occur. 

3.5. Explainability: Class Activation 

Mapping (CAM) 

The integration of Class Activation Mapping 

(CAM) improves the model's 

interpretability. CAM gives doctors visual 

feedback by identifying the areas of the 

input image that have the greatest influence 

on the model's decision-making process: 

Feature Map Activation: Feature maps are 

produced by the final convolutional layer of 

the feature extractor and are subsequently 

weighted based on their significance for the 

classification: 

𝐴𝑐(𝑥, 𝑦) = ∑  𝑘 𝑤𝑘
𝑐𝑓𝑘(𝑥, 𝑦)                                            

(11) 

where 𝑓𝑘(𝑥, 𝑦) is the activation of the 𝑘-th 

feature map at spatial location (𝑥, 𝑦), and 

𝑤𝑘
𝑐 is the weight associated with class 𝑐. 

Heatmap Development: The class activation 

map is generated by adding the weighted 

activations together. This map is then 

superimposed on the input image to create a 

heatmap that shows the important areas of 

the picture that are utilised for diagnosis. 

3.6. Addressing Data Imbalance: 

Synthetic Oversampling Techniques 

The Synthetic Minority Over-sampling 

Technique (SMOTE) is used to address the 

problem of imbalanced datasets, where 

specific classifications of skin diseases are 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

11 
      
 

under-represented. By generating synthetic 

samples through interpolation between pre-

existing minority class samples, SMOTE 

produces: 

𝑥new = 𝑥𝑖 + 𝜆(𝑥𝑗 − 𝑥𝑖)                                                     

(12) 

where 𝑥𝑖 and 𝑥𝑗 are samples from the 

minority class, and 𝜆 is a random value 

between 0 and 1. 

By increasing the representation of minority 

classes, this method improves the 

generalisation of the KNN model across all 

categories of skin diseases. 

3.7. Clinical Data Integration 

Alongside image-based features, clinical 

metadata 𝐶 (such as symptoms, patient 

history, and demographics) is integrated into 

the model. This multi-modal approach 

ensures that both visual features and clinical 

data are considered in the final 

classification, improving diagnostic 

accuracy. The combined feature set is 

represented as: 

𝐹 = 𝐹𝑖𝑚𝑔 ∪ 𝐶                                                        

(13) 

where 𝐹img  represents the image features, 

and 𝐶 is the clinical data. 

3.8. Skin Disease Classification and 

Evaluation 

The system's final output is the input image's 

classification into one of the pre-established 

categories for skin diseases. Standard 

classification metrics, such as accuracy, 

precision, recall, and F1-score, are used to 

assess the model's performance: 

Accuracy: 

 Accuracy =
𝑇𝑃+𝑇𝑁

𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁
                                                 

(14) 

Precision: 

 Precision =
𝑇𝑃

𝑇𝑃+𝐹𝑃
                                                       

(15) 

Recall: 

 Recall =
𝑇𝑃

𝑇𝑃+𝐹𝑁
                                                       

(16) 

F1-score: 

𝐹1 = 2 ⋅
 Precision ⋅ Recall 

 Precision + Recall 
                                                 

(17) 

where 𝑇𝑃, 𝑇𝑁, 𝐹𝑃, and 𝐹𝑁 represent true 

positives, true negatives, false positives, and 

false negatives, respectively. 

Through the CAM heatmaps, the system 

ensures that a diagnosis is comprehensible 

for physicians, giving them confidence in 

the outcomes and insight into the model's 

decision-making process. 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

12 
      
 

 

Figure 1: Accurate Skin Disease Detection Using K-Nearest Neighbors and CAM in IoMT-

Enabled Diagnostic Solutions. 

The flow of the suggested system for 

detecting skin diseases is shown in Figure 1. 

IoMT Image Acquisition is the first step, in 

which linked devices are used to take skin 

images. These pictures undergo 

preprocessing, which includes scaling, noise 

reduction, and contrast enhancement. The 

lesions are then separated using DF-U-Net 

segmentation. Texture and morphological 

features are recovered from the segmented 

images using feature extraction. These 

features are then used by the K-Nearest 

Neighbours (KNN) Classification algorithm 

to categorise the skin condition. Heatmaps 

indicating significant regions are produced 

by Class Activation Mapping (CAM), which 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

13 
      
 

improves interpretability. SMOTE is also 

used by the system to handle data 

imbalance. Ultimately, the approach 

produces a visual description and a precise 

classification of skin diseases. 

4. RESULT AND DISCUSSION 

K-Nearest Neighbours (KNN) and Class 

Activation Mapping (CAM) are combined in 

the proposed IoMT-enabled skin disease 

diagnosis system to obtain higher accuracy 

and interpretability. Preprocessing methods 

that greatly improve image quality and 

enable more accurate analysis include DF-

U-Net segmentation and CLAHE for 

contrast enhancement. Essential texture 

features are extracted with the help of the 

Gray-Level Co-occurrence Matrix (GLCM), 

and these features are critical to the KNN-

based classification of skin diseases. In order 

to balance bias and variance and improve the 

model's performance on minority skin 

disease classes, SMOTE resolves data 

imbalance, while cross-validation guarantees 

the selection of the ideal K-value. 

Class Activation Mapping (CAM) produces 

visual heatmaps that emphasise important 

regions of the image that affect the 

classification, which improves 

interpretability. With a 99.18% accuracy, 

99.30% sensitivity, and 99.16% precision, 

the system outperformed other approaches 

like XGBoost and EFF_D_SVM with 

remarkable results. High classification 

accuracy can only be achieved by include 

essential components, including DF-U-Net 

segmentation and CAM, whose removal 

results in significant performance decreases. 

This is demonstrated by the ablation study. 

Table 1: Comparison of the Proposed KNN + CAM System with Other Methods. 

Technique Accuracy (%) Sensitivity (%) Precision (%) F1-Score (%) FNR (%) 

Proposed KNN + 

CAM 

99.18 99.30 99.16 99.34 0.48 

EFF_D_SVM 

(Zhang et al., 

2023) [1] 

95.12 94.75 97.52 95.12 4.88 

XGBoost (Rezaee 

et al., 2022) [2] 

97.12 96.54 96.97 96.85 2.88 

KNN (Araaf and 

Nugroho, 2023) [9] 

92.60 91.30 92.00 91.90 7.40 

SVM (Oumarou, 

2023) [7] 

93.45 92.60 94.35 93.10 6.55 

The performance of the suggested KNN + 

CAM system is displayed in Table 1 in 

comparison to other methods cited in the 

literature. The suggested system outperforms 

existing techniques such as EFF_D_SVM 

(Zhang et al., 2023), XGBoost (Rezaee et 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

14 
      
 

al., 2022), and the conventional KNN 

algorithm (Araaf and Nugroho, 2023) with 

the maximum accuracy of 99.18% and the 

lowest False Negative Rate (FNR) of 0.48%. 

The table illustrates how the accuracy and 

interpretability of the suggested skin disease 

detection system are greatly enhanced by 

combining CAM with preprocessing 

methods like DF-U-Net segmentation and 

SMOTE for data balancing. 

 

Figure 2: Comparison of the Proposed KNN + CAM System with Other Methods. 

The suggested KNN + CAM system is 

contrasted with different machine learning 

methods, including XGBoost, 

EFF_D_SVM, and traditional KNN, in 

Figure 2. The suggested system performs 

better than the other approaches in terms of 

F1-score, accuracy, sensitivity, precision, 

and False Negative Rate (FNR), as shown in 

the figure. The suggested strategy achieves 

better diagnostic accuracy (99.18%) and 

interpretability by combining CAM with 

sophisticated preprocessing techniques, 

proving the value of a hybrid approach over 

more conventional methods for the diagnosis 

of skin diseases. 

Table 2: Ablation Study of the Proposed IoMT-Enabled Skin Disease Detection System. 

 

Component Removed Accuracy 

(%) 

Sensitivity 

(%) 

Precision 

(%) 

F1-Score 

(%) 

Full System (Proposed) 99.18 99.30 99.16 99.34 

Without DF-U-Net Segmentation 93.75 94.12 93.80 94.32 

99.18 99.3 99.16 99.34

95.12
94.75

97.52

95.12

97.12

96.54
96.97 96.85

92.6

91.3

92 91.9

93.45

92.6

94.35

93.1

86

88

90

92

94

96

98

100

Accuracy (%) Sensitivity (%) Precision (%) F1-Score (%)

P
er

ce
n

ta
g

e

Proposed KNN + CAM EFF_D_SVM XGBoost KNN SVM



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

15 
      
 

Without CAM 96.32 96.75 96.55 96.60 

Without SMOTE (Data Imbalance 

Handling) 

94.12 93.85 93.91 93.88 

Without Clinical Data 95.45 95.70 95.10 95.34 

Without GLCM Feature Extraction 90.34 90.10 89.80 89.95 

Without Preprocessing 88.45 87.65 88.20 87.90 

Without Hyperparameter Tuning (K) 92.10 91.50 91.90 91.70 

The impact of deleting particular parts from 

the suggested system is assessed in the 

ablation study Table 2. The system's 

accuracy drops significantly to 93.75% in 

the absence of DF-U-Net segmentation, 

underscoring the significance of accurate 

lesion detection. Interpretability is decreased 

when CAM is excluded, while performance 

on imbalanced datasets is adversely affected 

when SMOTE is removed. Accuracy is 

reduced when clinical data or GLCM feature 

extraction are excluded, highlighting the 

importance of multi-modal learning. 

Furthermore, there is a discernible 

performance reduction when preprocessing 

or hyperparameter adjustment (ideal K-

value) are removed, highlighting the 

importance of each component in improving 

system accuracy and generalisation. 

 

80

82

84

86

88

90

92

94

96

98

100

102

Accuracy (%) Sensitivity (%) Precision (%) F1-Score (%)

P
er

ce
n

ta
g

e 
(%

)

Full System (Proposed) Without DF-U-Net Segmentation

Without CAM Without SMOTE

Without Clinical Data Without GLCM Feature Extraction

Without Preprocessing Without Hyperparameter Tuning (K)



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

16 
      
 

Figure 3: Ablation Study of the Proposed IoMT-Enabled Skin Disease Detection System. 

The ablation research findings for the 

suggested IoMT-enabled skin disease 

detection system are shown in Figure 3. It 

illustrates how the system's performance is 

affected when important elements like 

CAM, SMOTE, and DF-U-Net 

segmentation are removed. The graphic 

shows how the accuracy and sensitivity of 

the model are greatly decreased in the 

absence of segmentation and data balancing 

approaches. This figure, which illustrates 

that the whole system with all components 

functions well across all metrics, highlights 

the significance of each component in 

maintaining high performance and accuracy. 

5. CONCLUSION 

The suggested IoMT-enabled diagnostic 

system provides a highly accurate and 

interpretable method for the diagnosis of 

skin diseases by combining K-Nearest 

Neighbours (KNN) and Class Activation 

Mapping (CAM). It outperforms 

conventional techniques with a low false-

negative rate of 0.48% and an accuracy of 

99.18%. Performance is further improved by 

preprocessing methods like SMOTE for data 

balance and DF-U-Net segmentation. This 

system is a trustworthy tool for healthcare 

decision-making because it blends 

sophisticated machine learning with 

interpretability. Future developments will 

involve growing the dataset for more 

generalisability across a range of skin 

conditions and integrating deep learning 

models, such EfficientNet, for better feature 

extraction. 

REFERENCES 

1. Surendar Rama Sitaraman (2021). 

Crow search optimization in AI-

powered smart healthcare: A novel 

approach to disease diagnosis. 

Journal of Current Science & 

Humanities, 9(1), 9-22. 

2. Zhang, J., Tan, X., Chen, W., Du, G., 

Fu, Q., Zhang, H., & Jiang, H. 

(2023). EFF_D_SVM: a robust 

multi-type brain tumor classification 

system. Frontiers in 

Neuroscience, 17, 1269100.  

3. Rezaee, K., Khosravi, M. R., Neshat, 

N., & Moghimi, M. K. (2022). Deep 

transfer learning-based fall detection 

approach using IoMT-enabled 

thermal imaging-assisted pervasive 

surveillance and big health 

data. Journal of Circuits, Systems 

and Computers, 31(12), 2240005. 

4. Pandey, R., Maurya, P., & Chiong, 

R. (Eds.). (2023). Data Modelling 

and Analytics for the Internet of 

Medical Things. CRC Press. 

5. Santana-Mancilla, P. C., Castrejón-

Mejía, O. E., Fajardo-Flores, S. B., 

& Anido-Rifón, L. E. (2023). 

Predicting abnormal respiratory 

patterns in older adults using 

supervised machine learning on 

Internet of medical things respiratory 

frequency data. Information, 14(12), 

625. 

6. Surendar Rama Sitaraman (2021). 

AI-Driven Healthcare Systems 

Enhanced by Advanced Data 

Analytics and Mobile Computing. 

International Journal of Information 



CTMJ | traditionalmedicinejournals.com Chinese Traditional Medicine Journal | 2024 | Vol7 |Issue3 

17 
      
 

Technology & Computer 

Engineering. 

 

7. Anikwe, C. V., Nweke, H. F., 

Ikegwu, A. C., Egwuonwu, C. A., 

Onu, F. U., Alo, U. R., & Teh, Y. W. 

(2022). Mobile and wearable sensors 

for data-driven health monitoring 

system: State-of-the-art and future 

prospect. Expert Systems with 

Applications, 202, 117362. 

8. Muneer, S., & Khan, S. (2023). 

Intelligent Disease Prediction System 

for Hepatitis C 

Patients. International 

Multidisciplinary Journal Of 

Science, Technology & 

Business, 2(04), 15-21. 

9. Oumarou, H. (2023). Detection and 

Classification of Bacterial Skin 

Infections Using K-Nearest 

Neighbors Algorithm. International 

Journal of Artificial Intelligence in 

Medical Issues, 1(2), 124-136. 

10. Esti, N. M. C., Dewi, M. R., Anry, 

N. K. N., & Pratikno, G. H. Z. (2023, 

September). Implementation K-

Nearest Neighbor Algorithm Skin 

Diseases Classifier. In 2023 

International Seminar on 

Application for Technology of 

Information and Communication 

(iSemantic) (pp. 502-507). IEEE. 

11. Araaf, M. A., & Nugroho, K. (2023). 

Comprehensive analysis and 

classification of skin diseases based 

on image texture features using K-

nearest neighbors algorithm. Journal 

of Computing Theories and 

Applications, 1(1), 31-40. 

12. Surendar Rama Sitaraman (2020). 

Optimizing Healthcare Data Streams 

Using Real-Time Big Data Analytics 

and AI Techniques. International 

Journal of Information Technology 

& Computer Engineering. 

International Journal of Engineering 

Research and Science & Technology.  

 

13. https://www.kaggle.com/code/dataf

an07/analysis-of-melanoma-

metadata-and-effnet-ensemble  

 

 

 

https://www.kaggle.com/code/datafan07/analysis-of-melanoma-metadata-and-effnet-ensemble
https://www.kaggle.com/code/datafan07/analysis-of-melanoma-metadata-and-effnet-ensemble
https://www.kaggle.com/code/datafan07/analysis-of-melanoma-metadata-and-effnet-ensemble

