









































49 J Global Clinical Engineering Vol.7 Issue 2: 2025

Received October 29, 2024, accepted May 15, 2025, date of publication June 17, 2025.

Case Study

Case Study: Augmented Reality Enabled Mental Health Chatbot

Subbaraj Pravin Kumar*, Akash Kumar, and Anusha Amba Prasanna

Biomedical Engineering, Sri Sivasubramaniya Nadar College of Engineering, Chennai, Tamil Nadu, India.

* Corresponding Author Email: pravinkumars@ssn.edu.in

ABSTRACT

Background and Objective: In recent years, there has been a growing demand for mental health support. This has led to 
a focus on providing personalized and continuous care. However, traditional mental health systems often have long wait times 
and limited support for engagements beyond clinical hours. The goal of this project is to create ARden, a digital companion using 
augmented reality, to help improve mental health for those in need. 

Material and Method: This study aims to fine-tune a Large Language Model with domain-specific knowledge, ensuring a 
personalized and intelligent companion—ARden. The chatbot is integrated with the AR companion using an Application Program 
Interface (API). The mixed reality companion is accessible via a mobile application, making care available without the additional 
hardware costs associated with head-mounted displays. 

Results: The development of ARden has introduced new possibilities for personalized and interactive mental health sup-
port. Early feedback suggests that the chatbot may help improve user engagement and satisfaction, supported by encouraging 
retention metrics. By combining augmented reality, large language models, and a character-based interface, ARden offers an 
approach that could contribute positively to mental health support. 

Conclusion: ARden aims to help users with emotional regulation during long wait times between mental health interventions, 
overcome communication barriers, and provide exercises and suggestions to improve mental health wellbeing. This approach 
offers a promising solution to existing mental health challenges and holds potential for further improvement and scalability. 

Keywords—Augmented reality, Large language model, Mood score.      

Copyright © 2025. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY): Creative Commons - 
Attribution 4.0 International - CC BY 4.0. The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright 
owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduc-
tion is permitted which does not comply with these terms.

http://www.globalce.org
http://globalce.org
http://globalce.org
https://creativecommons.org/licenses/by/4.0/
https://creativecommons.org/licenses/by/4.0/


P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

J Global Clinical Engineering Vol.7 Issue 2: 2025 50

INTRODUCTION

Mental health problems pose a critical public health 
burden, yet traditional solutions remain insufficient to 
address the growing demand. The advent of social media 
and increased isolation among the younger population 
have exacerbated mental health challenges for these 
individuals. According to NextStep Solutions, a leading 
provider of behavioral health software, 29% of the U.S. 
population experiences at least one form of mental illness.1 

Along with the factors that have led to an increase in 
mental health issues, there is also a severe shortage of 
psychiatrists, with only nine per 100,000 people. This 
deficit contributes to negative mental health outcomes, 
including an increased risk of suicide.1 

At present, individuals seek help from their family, 
social circles, the internet, and mental health profession-
als to overcome these issues. However, several barriers 
limit the effectiveness of current mental health systems, 
including social stigma and a shortage of mental health 
professionals. 

Currently, most of the mental health help that individu-
als receive is in the form of cognitive behavioral therapy 
and medication. Cognitive behavioral therapy (CBT) is 
based on the idea that professionals are trained to help 
individuals overcome the issues they are facing and 
enable them to tackle problems on their own by using 
structured systems. However, due to understaffing, the 
current system struggles with increased costs, long wait 
times, and other challenges. Despite its shortcomings, 
seeking professional help remains the best option for 
those facing mental health challenges. 

Recent studies have explored the potential of chatbots 
to improve mental health outcomes. Denecke et al.1 ex-
amined two prominent AI-driven mental health chatbots, 
WYSA and SERMO. Wysa Inc., headquartered in the USA, 
is an everyday mental health application, and Sermo, also 
headquartered in the USA, is a social platform for physicians 
to collaborate and stay informed. WYSA detects negative 
moods and integrates features like depression assess-
ments and meditation exercises, while SERMO addresses 
psychological impairments using Cognitive Behavioral 
Therapy (CBT) techniques. Quantitative analysis revealed 

that frequent users of WYSA experienced greater mood 
improvements than occasional users. Experts acknowl-
edged that SERMO is well-suited for patients struggling 
with face-to-face communication, although challenges 
remain, including issues with data retention, dataset 
generation, and the inability of AI systems to handle 
emergencies effectively. 

Potts et al.2 addressed the lack of accessible mental 
health services in rural areas by introducing a multilingual 
chatbot, ChatPal. It was developed by the academic con-
sortium led by Ulster University (UK), which collaborated 
with partners in Ireland, Scotland, Sweden and Finland. 
Funded by the Northern Periphery and Arctic (NPA) 
Programme, ChatPal aimed to provide support in English, 
Scottish Gaelic, Swedish, and Finnish. The study employed 
a single-arm pre-post intervention design, enrolling par-
ticipants from rural areas to use ChatPal over 12 weeks, 
with well-being measured via scales like SWEMWBS. 
ChatPal, developed with Rasa (backend) and PhoneGap 
(frontend), features mood logging and mindfulness exer-
cises. While ChatPal is seen as a complementary tool for 
mental health services, the study calls for further research 
to confirm its effectiveness. The chatbot’s multilingual 
capabilities improve accessibility, but technical issues 
remain, particularly with integration and functionality. 

Social desirability and social support are factors that 
contribute to positive mental health outcomes. Similarly, a 
need to belong, which is heightened among young adults, 
is associated with negative outcomes. Thus, looking at the 
current landscape of mental health issues, it is clear that 
providing emotional support through interventions can 
help people, but is no replacement for a true emotional 
connection with others. It can serve to help people deal 
with issues in a healthier way, leading to better outcomes. 

Another study aimed to evaluate the effectiveness 
of ChatGPT in providing mental health support, with a 
particular focus on anxiety and depression. The primary 
objective was to assess the quality of responses generated 
by ChatGPT to user queries related to these mental health 
conditions. The study specifically analyzed the model's 
responses to three queries: two relating to anxiety man-
agement and medication, and a third regarding alterna-
tive treatment options. Moreover, the study examined the 



51 J Global Clinical Engineering Vol.7 Issue 2: 2025

P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

consistency and reliability of ChatGPT’s responses across 
successive interactions.

It was found that ChatGPT did not provide information 
about medication.3 Some advantages include its ability 
to offer personalized advice based on a person's history, 
improved accessibility in remote locations, and lower 
costs compared to traditional therapy. However, the infor-
mation must be cross-checked with professionals due to 
some inconsistencies, and it cannot substitute for mental 
health care. The model also cannot provide prescriptions. 
ChatGPT’s responses to prompts can be inappropriate, 
possibly due to the type of questions being asked. Thus, 
while ChatGPT is useful, the model developed for the user 
must address its shortcomings in some way.

Yang et al.4 investigated the capabilities of large language 
models (LLMs) within the healthcare domain, focusing 
on both their potential applications and inherent limita-
tions. The primary goal was to assess the effectiveness of 
general LLMs in healthcare settings and to identify areas 
where domain-specific models could offer improved per-
formance. General-purpose LLMs often lack the specialized 
knowledge required for healthcare applications due to the 
disparity between the general text used in their training 
and the professional, domain-specific content needed for 
clinical use. The study highlighted the promise of domain-
specific LLMs. For instance, BioBERT was developed by 
Korea University and trained on PubMed data. Another 
example, SCIBERT, was created by the Allen Institute 
for AI and was trained on broad scientific texts from 
Semantic Scholar. Similarly, PubMedBERT, by Microsoft 
Research, was specifically trained on PubMed abstracts. 
These models are all based on the BERT architecture and 
require significant computational resources for opera-
tion. Despite their promise, the study acknowledged the 
challenges that remain in their clinical implementation.

The performance of domain-specific LLMs was found 
to be superior compared to general models, particularly 
in patient interactions. A tailored model called ChatDoc-
tor, which is a fine-tuned large language model based on 
Llama and trained on 100,000 real-world patient-doctor 
dialogues from an online consultation platform, and sup-
ported by an NIH grant. It demonstrated enhanced efficacy 
in clinical settings. However, the study also identified 

significant challenges in deploying LLMs in healthcare, 
particularly concerning data integrity, interpretability, and 
the high costs associated with developing these models. 
The integration of LLMs into clinical practice as supple-
mentary tools was explored, emphasizing the need for 
improvements in task optimization and conversational 
assistance.

The challenges related to interpretability, data limi-
tations, and ethical considerations must be addressed 
to fully realize the potential of LLMs in clinical practice. 
The research suggests that future efforts should focus 
on optimizing these models for specific tasks, improving 
data diversity, and ensuring the accuracy and reliability of 
the content generated by LLMs, particularly when cross-
referenced with professional expertise.

User retention is also reported to be a critical factor, 
emphasizing the importance of highly engaging interac-
tions with chatbots. It was demonstrated that inadequate 
retention rates often stem from a lack of personalization, 
which impedes the effectiveness of mental health apps. 
To address these challenges, the paper proposed several 
solutions, including the personalization of chatbots by 
utilizing the user’s previous conversations. Additionally, 
incorporating peer communication methods was found 
to enhance both engagement and effectiveness. The de-
veloped app should adapt based on user feedback, with 
the overall goal of creating a more user-centric, adaptable, 
and effective platform.

In another study, data from the chatbot interactions, 
including session details and mood logs, were analyzed to 
extract features such as tenure, mood logging frequency, 
and conversation interactions. K-means clustering was 
employed to categorize users into three groups: abandon-
ing, frequent transient, and sporadic users. This analysis 
compared user behaviors, engagement, and retention 
metrics with those of other mental health apps. The 
study emphasizes the importance of high engagement 
and retention metrics and the need for personalized user 
experiences.5

The effectiveness of the chatbot heavily depends on 
the underlying language model that powers it. Llama 2, 
an advanced open-source language model from Meta AI, 
can generate text similar to that of a human and is useful 



P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

J Global Clinical Engineering Vol.7 Issue 2: 2025 52

Mental health professionals use screening question-
naires (SQs) to identify symptom areas that require fur-
ther exploration. Regular screening can enable the early 
identification of individuals in high-stress professions 

who may require mental health support. Data indicate 
that a significant percentage of public safety personnel 
screen positive for at least one mental health disorder, 
highlighting the advantages of frequent screening.9 

Integrating LLMs into chatbots could enhance their 
ability to provide tailored support, especially when 
fine-tuned for specific screening tasks within high-stress 
populations.

Incorporating Virtual Reality (VR) and Augmented 
Reality (AR) into mental health interventions offers a 
transformative approach to enhancing user engage-
ment and interaction. They can improve access to and 
availability of therapy due to their personalized nature. 
Proper training for mental health professionals, rigorous 
scientific research, and strict adherence to data privacy 
and ethical guidelines are essential for the responsible 
use of mental health apps, making them more engaging, 
targeted, and therapeutically effective.

Integrating Augmented Reality (AR) into chatbot 
platforms represents a promising advancement in mental 
health care. Current mental health chatbots, while offering 
useful features, have limitations. One major bottleneck 
is the need to ensure data privacy, especially since chat-
bots that provide personalized suggestions must store 
previous user interactions.2 Ensuring that user data is 
protected from unauthorized use is crucial for trust and 
widespread adoption.

Another challenge is the need for relevant content and 
fine-tuning large language models to provide helpful and 
contextually appropriate responses. It must converse 
with the user in a manner that is genuinely helpful to 
them. Fine-tuning the model requires resources, and 
the model needs continuous updates to stay relevant.3 A 
drawback is that the way the model arrives at decisions 
is not explained clearly.

User retention is another significant challenge for 
chatbot applications. Generalizations about user behavior 
have led to the development of different user archetypes.4 
This information about how users interact with the chat-
bot can be used to improve the retention metrics of the 
application we build.

for chatbots, translation, content production, and other 
applications. Llama 2 offers notable advantages in versatil-
ity and adaptability when fine-tuned for specific domains, 
addressing limitations observed in previous models.6 

Roumeliotis et al.6 investigated the challenges and 
opportunities faced by developers when deploying and 
fine-tuning Llama 2, with the hypothesis that the open-
source nature of Llama 2 facilitates faster development 
compared to closed-source models. Early adopters’ 
experiences in deploying and fine-tuning Llama 2 were 
observed over a 10-day period, with particular attention 
given to the medical domain, a primary area of interest 
for fine-tuning efforts.

Data on model deployment, fine-tuning, and other 
relevant factors were gathered during this period. Textual 
data was then processed through keyword identification, 
K-means clustering, and word cloud visualization. The 
resulting analysis reveals that Llama 2 can be seamlessly 
deployed and fine-tuned to the domain-specific require-
ments of various industries, thereby addressing challenges 
encountered with earlier models.

Yang et al.7 compared four Large Language Models 
(LLMs) for mental health analysis, focusing on the effec-
tiveness of prompting strategies such as Chain of Thought 
prompting, emotion-enhanced prompts, and few-shot 
learning. The findings emphasized the importance of 
domain-specific fine-tuning for improved results in build-
ing mental health solutions.

Regular large language models often fall short in 
specialized areas like medicine, where domain-specific 
knowledge is crucial.8 The authors proposed PMC-Llama, 
an open-source language model specifically tailored for 
medical applications.8 They systematically analyzed the 
process of adapting a general-purpose LLM to the medical 
domain by integrating 4.8 million biomedical academic 
papers and 30,000 medical instructional materials, and 
extensively fine-tuned it for compliance with the domain-
specific knowledge base.



53 J Global Clinical Engineering Vol.7 Issue 2: 2025

P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

For chatbots to be effective for emotional support, they 
must understand how users feel and react empathetically. 
This is possible by using libraries with words and the emo-
tions associated with those words. However, chatbots are 
only as good as their prompts, and the sentiment and emo-
tion lexicons used for emotion-enhanced prompts suffer 
from annotation bias and limited vocabulary, which may 
not reflect the evolving language used in recent datasets.4 

Recent advancements in training large language models 
across multiple languages have increased accessibility and 
improved the generalizability of studies.5 Although this is 
important, the focus of this paper is on developing a ro-
bust mental health chatbot primarily for English-speaking 
users, as they represent the majority of current users. 

Given the limitations of existing mental health chatbots 
and the recent advancements in technology, this paper 
proposes the development of a chatbot using the Llama 
2 model integrated with Augmented Reality technology. 
The proposed chatbot will prioritize high user retention, 
accuracy, privacy, and address current chatbot limitations 
while incorporating the additional functionalities discussed.

This study proposes an Augmented Reality-enabled 
Mental Health Chatbot that can provide supplementary 
support between visits to mental health professionals. 
Although chatbots cannot replace traditional therapy, 
they can offer continuous mental health support, helping 
individuals declutter their thoughts and providing acces-
sible care at any time. 

The aim was to analyze user engagement with a mental 
health chatbot, focusing on its potential to improve user 
retention through interactive and personalized experiences. 
User retention was examined to identify challenges and 
optimize engagement by understanding different user 
archetypes. The impact of the chatbot on users’ mental 
health needs to be monitored over time when used along-
side professional medical guidance.

The chatbot was developed as an app, making it ac-
cessible to a broader audience with user-friendliness. 
This study used LLM models that enable personalization, 
which is critical for effective content delivery. Various 

personalization techniques, such as retaining the memory 
of previous conversations, were implemented to create a 
more tailored and continuous user experience.

Retaining the memory of previous conversations helped 
continue interactions from the last episode, rather than 
starting from the beginning each time. 

METHODOLOGY

Workflow

The large language model (LLM) is customized with 
extensive medical literature and fine-tuned for optimal 
effectiveness. The application processes auditory input 
to provide information to the LLM. Additionally, the ap-
plication features a character that users interact with, 
designed to appear friendly and empathetic. This compan-
ion, integrated with the LLM, facilitates user interaction 
and contributes to mental health improvement through 
the application’s functionalities.

Data 

A mental health chatbot requires vast amounts of data 
to provide accurate and reliable solutions to individuals’ 
mental health challenges. Developing a comprehensive 
database that integrates both local knowledge and online 
resources is essential.7 This database also contains research 
articles tailored to the individual’s specific needs, such as 
those addressing emotional support, depression, anxiety 
disorders, and eating disorders. General information re-
garding interventions can also be obtained from reputable 
online sources, complementing the personalized data.

Profiling 

Mental health professionals initially assess and profile 
individuals based on their diagnosis, which is then fed as 
input to the model (Figure 1). This profiling enhances the 
personalization of the chatbot. Users are prompted to select 
a broad category that aligns with their experiences, and 
the application provides relevant information based on 
both the user’s selection and the therapist’s assessment



P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

J Global Clinical Engineering Vol.7 Issue 2: 2025 54

.

This work aims to understand individual perspectives 
and provide relief from mental health conditions, such 
as stress, anxiety, and depression, through personal-
ized solutions. To effectively address the needs of users 
experiencing mental health issues, collaboration with 
professionals is essential. Through multiple sessions 
with a counseling psychologist, they assess a person’s 
characteristics, drawing upon their experience and intu-
ition developed through years of study and practice. The 
chatbot can then create a tailored solution or personalized 
interaction method for each individual, taking profes-
sional input into account. For instance, if a person tends 
to respond only to non-confrontational communication, 
the chatbot will recognize this and deliver information 
in a compassionate, non-threatening manner to ensure 
the individual is receptive.

A questionnaire is developed, as shown in Figure 2, 
based on input from the psychologist, with its design 
undergoing thorough consultation and multiple levels 
of review by the therapists. This personalized profiling 
ensures that users receive appropriate, targeted support, 
rather than generic interventions.

Mental Health Chatbot

Large Language Model

The process begins by uploading appropriate resources 
as PDF files, followed by the application of a text extrac-
tion algorithm to retrieve the content while removing       

any non-text elements (Figure 3).8 The extracted text 
is then segmented into manageable pieces to facilitate 
meaningful embeddings. Llama 2, implemented via the 
Langchain framework, is employed to generate embeddings 
that capture the semantic essence of each text fragment. 
These embeddings are stored and managed in Pinecone, 
a cloud-based vector store. This fully managed service 
handles hardware infrastructure required for efficiently 
storing and searching vector data. Once the embeddings 
are uploaded to Pinecone’s cloud storage, natural language 
queries can be processed. Pinecone performs a similarity 
search by converting queries into text representations 
and generating embeddings via Langchain Llama 2. This 
search identifies the most similar embeddings, retriev-
ing the corresponding text fragments. These fragments 
are then combined to form a cohesive natural language 
response, enabling smooth and effective user interaction 
in question-and-answer scenarios.

FIGURE 1. Workflow of AR Companion (ARden) assisted per-
sonalized therapy.

FIGURE 2. Questionnaire.



55 J Global Clinical Engineering Vol.7 Issue 2: 2025

P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

Model Selection

Llama 2 was selected for this preliminary study due 
to several compelling factors. Its open-source framework 
supports accessibility and collaborative development, 
offering a parameter range from 7 billion to 70 billion. 
Additionally, Llama 2 distinguishes itself through its speed, 
which outperforms earlier models. This is particularly 
advantageous for time-sensitive tasks and applications 
requiring rapid processing, like mental health chatbots. 
Moreover, Llama 2 offers comprehensive documentation 
and a supportive open-source community, facilitating 
its integration. As large language models evolve rapidly, 
future iterations of this work will consider adopting more 
advanced models.9 

Integration with Chatbot

Firstly, a Unity project was set up with the necessary 
dependencies installed to enable the companion to be 
deployed into a mobile application. Inworld AI is an en-
gine that was used to create a character prefab to import 
into Unity. To begin with, the Inworld AI software had to 
be downloaded and interfaced with the Unity platform. 
Then, after the avatar was generated using Ready Player 
Me, a cross-game avatar platform that enables avatar 
creation and seamless integration into other platforms, 
its characteristics were customized using the Inworld 
AI portal. The API keys were then configured within the 
Unity project to import the character into the environ-
ment. The built-in speech of the character was replaced 
with the mental health chatbot. Visual customizations 
from Ready Player Me were refined, and Inworld AI’s 
tools were used to adjust baseline emotional expressions 
and idle animations to suit a mental health companion.

The imported character in the project had to be po-
sitioned, rotated, and scaled to the correct proportions 
relative to the room. This process required trial and error, 
and it had to be made such that it aligned with the real-
world surfaces and surroundings based on the camera 
position. The avatar was also fitted with a script to make 
it move wherever the user desired by just clicking on the 
spot. This was done to ensure the companion was placed 
in the position where the user was most comfortable. This 
level of personalization was aimed at giving the user the 
best experience possible. 

Inworld AI, a character engine developed by TheGist, 
Inc. (dba Inworld AI, New York, NY, USA, with its platform 
publicly launched c.2022 and continuously updated), is 
used to create non-playable characters in games using 
AI, natural language processing, emotional simulation, 
and behavior modeling. It generates expressions and 
facial features based on the character’s behavior using AI 
models that mimic human gestures.10 The avatars were 
also modified in this study. This behavior is determined 
by the information we feed it. Characteristics such as 
anger, confidence, and aggressiveness, for example, can 
be changed using the user interaction tools. 

These characteristics influence how the avatars com-
municate and interact with the user. This can be brought 
to the user by our custom mental health chatbot, which 
bypasses the default conversational settings.

The first step was to create the chatbot, which has 
already been described. Post-chatbot creation, it is to be 
interfaced with the character instead of the in-built GPT-3 
model. Emotion detection is another important aspect of 

FIGURE 3. Knowledge base construction.



P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

J Global Clinical Engineering Vol.7 Issue 2: 2025 56

this application that needs to be improved upon in the 
next iterations of it.11

User Interface (UI)

The initial version of the UI had a canvas with buttons 
to allow the user to navigate between different function-
alities of the application. However, this was removed to 
truly make the application even more user-friendly and 
non-frustrating. The idea of the avatar companion fulfill-
ing all of the user’s commands and the avatar acting as 
the interface was more appropriate. Thus, the UI was 
changed to make the companion-user relationship the 
most important aspect of the project. This also enables 
future iterations of the application to have more interest-
ing use cases and functionalities.12

Natural Interactions

Talking with the character mimics human conversation. 
This is an intuitive approach that lowers the cognitive 
load for the user. This lower barrier to entry can be the 
difference between the app being used or not.

Convenience

 As the interface is completely voice-based, people with 
disabilities, such as motor or vision impairments, can use 
the application. Going forward, it can use the information 
from the user to guide them to take medication on time. 

Efficiency

The commands are much faster when voice-based 
compared to navigating through menus and typing the 
queries, as with conventional chatbots. 

Personalization

As the chatbot learns from previous interactions with 
the user, this brings a level of personalization to the user 
interface that is just not possible through hard-coded 
menus. The chatbot has the ability to give personalized 
information in less time. It is hypothesized that, as the 
relationship between the user and ARden grows, the ap-
plication will achieve higher retention and usage rates. 
This is much more than what a traditional interface can do.

RESULTS

Recommendation From Professionals
The AI companion was developed in consultation 

with mental health professionals to ensure the ethical 
integrity of the application while validating the respon-
sible AI training, clinical relevance, and accuracy of their 
responses. The application developed primarily comple-
ments the therapeutic practices, addressing the gap in 
mental health delivery systems.13 In alignment with this 
approach, the proposed idea of this AR-enabled chatbot 
application was taken to a mental health professional at 
the National Institute for Empowerment of Persons with 
Multiple Disabilities (NIEPMD) in India for initial valida-
tion. The key suggestions that were implemented include: 
ease of use, efficient interaction without too many menus, 
and personalization. 

The other recommendation was to target general 
mental health needs, thereby broadening the scope of 
the application. This further enhanced the efficacy of the 
large language model by allowing it to specialize in high-
demand and specific areas of mental health. Subsequent 
steps focused on niche areas such as stress management, 
anxiety, and depression, to name a few.

Once a basic prototype was developed, it underwent 
further evaluation for feedback on clinical efficacy. The 
user base was defined in the review process of the pro-
totype. The prototype was intended to be prescribed 
to selected users based on their psychological profiles, 
following consultation with a certified mental health pro-
fessional. Consequently, the chatbot should be deployed 
as part of a hybrid healthcare model, ensuring that the 
mental health professional remains actively involved in 
the patient care loop.

Patient confidentiality was a major concern. Based on 
these inputs and further research, it was evident that any 
solution in this domain must strictly adhere to data privacy 
regulations.14 While our study has not yet encountered 
situations that raise privacy concerns, future research and 
similar product developments will need to ensure robust 
measures that protect patient data.

Parameter Tuning
As outlined in Figure 4, the mental health chatbot of-

fers a range of features that are designed to be beneficial 



57 J Global Clinical Engineering Vol.7 Issue 2: 2025

P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

to patients. To achieve this, various parameters are 
optimized to define the state-of-the-art large language 
model-based chatbot. Fine-tuning is accomplished by 
adjusting parameters such as temperature, maximum 
generation length, and sampling, among others.15 Rigor-
ous evaluations of the developed model are performed 
to mitigate risks and streamline responses to align with 
the user’s needs and expectations. 

The paradigm of the response is facilitated by the 
availability of these parameters. Temperature is a variable 
that tells us how incidental the output or the response 
generated is.16 The values can be interpreted as follows: 
if the temperature values are low, the model is more de-
terministic. This is ideal for maintaining consistent and 
reassuring conversations in mental health settings. Higher 
values may introduce more variability, which may not be 

required for this purpose. Due to this, the temperature 
was originally set to 0.5.17 

Tokens with a combined likelihood exceeding a threshold 
(p) are examined in top-p sampling, which is sometimes 
referred to as nucleus sampling. This restricts the token 
selection process by regularly modifying the queries ac-
cording to their probabilities. This approach preserves 
focus on the majority of tokens while ensuring diversity. 

As shown in Table 1, the maximum length of the re-
sponse generated is set to be 512 words. Seeing as shorter 
responses can keep the conversations focused but might 
lead to leaving out details, this is an area with tradeoffs. 
Other parameters, such as learning rate and beam search 
width, also play a crucial role in making the model dynamic 
and user-friendly. The existing learning rate of the large 
language model is 0.0002.18

FIGURE 4. Features of AR companion.



P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

J Global Clinical Engineering Vol.7 Issue 2: 2025 58

the entire conversation, the companion constantly exhibits 
friendly body language. 

As seen in Figure 5, the character interacts with the 
user with compassion and empathy, which is crucial when 
dealing with sensitive topics. In Figure 6, the chatbot asks 
the user to rate their feelings on a scale of 1 to 10, which 
lets the chatbot deal with the situation differently based 
on the user’s response. When the situation is particularly 
difficult, the user is directed to consult with the therapist. 

In Figure 7, the user is being suggested exercises to 
follow in order to feel better. In general, the exercises that 
therapists suggest to their patients can be reinforced in 
this app in order to ensure proper completion. Also, the 
app’s interactive nature immerses the user in the exer-
cise, leading to improved outcomes. In summary, the ap-
plication interacts with the user in a compassionate way 
to provide insights and aims to improve mental health 
outcomes. An integrated mechanism for analyzing and 

A model that is fine-tuned with relevant mental health 
data and uses emotion tracking tools to adjust the model’s 
tone and word choice would be ideal for this purpose. 
Setting parameters that encourage the chatbot to ask 
questions and take pauses would ensure user engage-
ment, thereby enhancing interaction.

The recent Llama 3 model launched by Meta represents 
a more advanced and efficient iteration of large language 
models, offering significant potential for future applica-
tions in the development of mental health chatbots. The 
following comparisons (Table 2) outline the key improve-
ments in the model, as indicated by Meta’s advancements.

Response of the chatbot
The chatbot shows compassion towards the user by first 

validating what the user feels and gently suggesting what 
the user could do to improve their situation. Throughout 

TABLE 1. Hyperparameters of Llama 2.

Parameters Value

Learning Rate 0.0002
Response Speed 120–250 s

Chunk Size 512

Temperature 0.5
Sampling Top P or nucleus sampling

Software Development Kit Boto3(AWS SDK for Python)

TABLE 2.  Comparison of Llama 2 and Llama 3.

Feature Llama 2 Llama 3

Training Data Trained on around 
2.2 trillion tokens

Trained on 
approximately 15 

trillion tokens

Model Sizes
Released in 7B, 13B, 
and 70B parameter 

sizes

Available in 8B and 
70B parameter 

versions
Context 
Window

Supports up to 4,096 
tokens

Supports up to 8,192 
tokens

Performance Better performance 
over Llama 2

Outperforms Llama 3 
across all benchmarks

FIGURE 5. Sample response for depression.

FIGURE 6. Sample response for anxiety.



59 J Global Clinical Engineering Vol.7 Issue 2: 2025

P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

presenting emotions detected from the users’ interac-
tions with the chatbot proved to be useful. Throughout 
these interactions, prominent emotional states such as 
happiness, sadness, and anger could be identified to help 
understand the mental state of the user. It was evident 
that this analysis would help the user track their mood. 

A formal study is planned to evaluate the therapeutic 
potential of the app. Participants’ moods will be assessed 
over a 30-day period using daily self-report scales, with ad-
ditional objective measures, such as physiological indicators 
or behavioral assessments, potentially incorporated. Data 
analysis will compare pre- and post-study mood scores, 
along with any improvements reported by participants. 
To further validate these findings, future research could 
focus on longitudinal studies and randomized controlled 
trials involving individuals diagnosed with mental health 
conditions, aiming to rigorously quantify the chatbot’s 
impact on mental health outcomes.

Testing Feedback
The application was tested for user interaction (UI) 

experiences, and some bugs were observed within the 
app, such as the character moving on its own at times, the 
app crashing during longer interactions, and the character 
not always being anchored in the environment correctly. 
In terms of functionality, it was suggested to add more 
exercises and activities, like guided meditations. 

Clinician Feedback
The designed application was evaluated by two differ-

ent psychiatrists. They suggested further work on adding 
more evidence-based exercises and ensuring the clinical 
accuracy of the information provided. They also wanted 
more features in the app so that clinicians can, with pa-
tient consent, access summaries of patients’ interactions 
or mood trends to better inform therapy sessions. Apart 
from this, they had questions regarding crisis manage-
ment protocols within the app and the specifications of 
data privacy and security for sensitive user information. 
Along with these important considerations, the overall 
feedback was encouraging, with psychiatrists recogniz-
ing the potential of Arden to support individuals between 
interventions and overcome communication barriers.19

CONCLUSION AND FUTURE WORK
As discussed, the application was developed after 

analyzing the capabilities and limitations of existing solu-
tions. It was determined that increased user engagement 
is essential for the application to be truly beneficial. With 
the advancement of AR technology and its capabilities, the 
study focused on designing a system that integrates both 
large language models (LLMs) and augmented reality (AR). 

The companion application was successfully imple-
mented using Unity and deployed on mobile platforms. The 
constructed avatar was equipped with an API interfacing 
with the LLM, which functions as a chatbot. This chatbot 
was fine-tuned with relevant medical literature to provide 
accurate and useful information to the user. Thus, the 
study produced an initial design of an AR-enabled avatar, 
equipped with an LLM, for emotional regulation purposes. 

FIGURE 7. Sample exercise.



P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

J Global Clinical Engineering Vol.7 Issue 2: 2025 60

AUTHOR CONTRIBUTIONS
All authors contributed equally to this work.

FUNDING
Not applicable.

DATA AVAILABILITY STATEMENT
Not applicable.

CONFLICTS OF INTEREST
The authors declare they have no competing interests.

ETHICS APPROVAL AND CONSENT TO PARTICIPATE 
Not applicable.

CONSENT FOR PUBLICATION
Not applicable.

FURTHER DISCLOSURE
Not applicable. 

REFERENCES
1. Denecke, K., Abd-Alrazaq, A., Househ, M. Artificial Intel-

ligence for Chatbots in Mental Health: Opportunities and 
Challenges. In Multiple perspectives on artificial intelligence in 
healthcare: Opportunities and challenges. Househ, M., Borycki, 
E., Kushniruk A. eds. Springer Nature, Switzerland; 2021; 
p.115–128; https://doi.org/10.1007/978-3-030-67303-1_10.

2. Potts, C., Lindström, F., Bond, R., et al. A multilingual digital 
mental health and well-being chatbot (ChatPal): pre-
post multicenter intervention study. J Med Internet Res. 
2023;25:e43051. https://doi.org/10.2196/43051. 

3. Farhat, F. ChatGPT as a complementary mental health resource: 
A boon or a bane. Ann Biomed Eng. 2024;52(5):1111–1114. 
https://doi.org/10.1007/s10439-023-03326-7.

4. Yang, R., Tan, T.F., Lu, W., et al. Large language models in health 
care: development, applications, and challenges. Health Care 
Sci. 2023;2(4):255–263. https://doi.org/10.1002/hcs2.61.

5. Booth F., Potts, C., Bond, R., et al. A mental health and 
well-being chatbot: user event log analysis. JMIR Mhealth 
Uhealth. 2023;11:e43052. https://doi.org/10.2196/43052.

The design developed in this study serves as a refer-
ence for future advancements in the field. Designing an 
effective mental health chatbot involves several critical 
aspects, including, but not limited to, tracking the mood 
scores of users,20 enhancing the LLM’s capabilities,21 
maintaining data privacy, and ensuring the chatbot retains 
context from previous interactions with the user.

Mood scores are an effective method for monitoring 
the mood and mental health of a person over a period of 
time. This approach can be validated with the assistance of 
mental health professionals.22 The system can offer more 
personalized support based on mood scores.

Collaboration with mental health professionals has 
offered important insights into the everyday challenges 
patients encounter. It is essential that products in this field 
are developed and tested alongside these professionals to 
ensure they effectively address real-world needs, particu-
larly by improving the mental health of young adults.23

This study was conducted prior to the release of Llama 
3, GPT-4, and other newer models. With further advance-
ments in large language models and emotional recognition, 
future chatbots are expected to exhibit more human-like 
qualities and be capable of taking in multimodal inputs 
from users more effectively.

Throughout the development of this application, a 
few key findings emerged regarding emotion detection. 
It became apparent that an application of this sort must 
take into account the following for the emotion detection 
capabilities: the user's speech and voice tone patterns, 
facial expression changes, linguistic cues, contextual 
awareness of the model, and multimodal data integration 
capabilities. These are all areas where humans naturally 
excel and are currently superior compared to models.

Real-time emotion detection remains a challenging 
task, which requires further research. Taking emotion 
recognition capabilities alongside the other functional 
abilities of the application previously discussed, it is 
clear that further research is required for applications 
of this kind before they can be effectively deployed. We 
anticipate that, with further technological advancements, 
this design of an AR-enabled mental health chatbot will 
contribute to improved mental health support.

https://doi.org/10.1007/978-3-030-67303-1_10
https://doi.org/10.2196/43051
https://doi.org/10.1007/s10439-023-03326-7
https://doi.org/10.1002/hcs2.61
 https://doi.org/10.2196/43052


61 J Global Clinical Engineering Vol.7 Issue 2: 2025

P. Kumar, A. Kumar, Prasanna: Case Study: Augmented Reality Enabled Mental Health Chatbot

14. May, R., Denecke, K. Security, privacy, and healthcare-related 
conversational agents: a scoping review. Inform Health Soc 
Care. 2022;47(2):194–210. https://doi.org/10.1080/175
38157.2021.1983578.

15. Sahoo, S.S., Plasek, J.M., Xu, H., et al. Large language models 
for biomedicine: foundations, opportunities, challenges, and 
best practices. J Am Med Inform Assoc. 2024;31(9):2114–2124. 
https://doi.org/10.1093/jamia/ocae074.

16. Boucher, E.M., Harake, N.R., Ward, H.E., et al. Artificially 
intelligent chatbots in digital mental health interventions: 
a review. Expert Rev Med Devices. 2021;18(sup1):37–49. 
https://doi.org/10.1080/17434440.2021.2013200.

17. Touvron, H., Martin, L., Stone, K., et al. Llama 2: Open 
foundation and fine-tuned chat models. arXiv preprint 
arXiv:2307.09288. 2023. 

18. Zhang, R., Han, J., Liu, C, et al. LLaMA-adapter: Efficient 
fine-tuning of large language models with zero-initialized 
attention. In the Twelfth International Conference on Learning 
Representations, Vienna Austria, 7–11 May, 2024. Available 
Online: https://openreview.net/pdf?id=d4UiXAHN2W.

19. Carlson, C. Virtual and augmented simulations in mental 
health. Curr Psychiatry Rep. 2023;25(9):365–371. https://
doi.org/10.1007/s11920-023-01438-4.

20. Hsu, M.C. The construction of critical factors for success-
fully introducing chatbots into mental health services in 
the army: Using a hybrid MCDM approach. Sustainability. 
2023;15(10):7905. https://doi.org/10.3390/su15107905.

21. Dosovitskiy, G. and Bunge, E. Development of a chatbot for 
depression: adolescent perceptions and recommendations. 
Child Adolesc Ment Health. 2023;28(1):124–127. https://
doi.org/10.1111/camh.12627.

22. Furukawa, T.A. Assessment of mood: guides for clini-
cians. J Psychosom Res. 2010;68(6):581–589. https://doi.
org/10.1016/j.jpsychores.2009.05.003.

23. Berryman, C., Ferguson, C.J., Negy, C. Social media use and mental 
health among young adults. Psychiatry Q. 2018;89(2):307–314. 
https://doi.org/10.1007/s11126-017-9535-6.

6. Roumeliotis, K., Tselikas, N., Nasiopoulos, D. Llama 2: Early 
adopters’ utilization of Meta's new open-source pretrained 
model. Preprints. 2023. https://doi.org/10.20944/pre-
prints202307.2142.v1.

7. Yang, K., Ji, S., Zhang, T., et al. Towards interpretable mental 
health analysis with large language models. In Proceedings 
of the 2023 Conference on Empirical Methods in Natural 
Language Processing,  Singapore, 6-10 December, 2023. 
https://doi.org/10.18653/v1/2023.emnlp-main.370.

8. Wu, C., Lin, W., Zhang, X., et al. PMC-LLaMA: toward building 
open-source language models for medicine. J Am Med Inform 
Assoc. 2024;31(9):1833–1843. https://doi.org/10.1093/
jamia/ocae045.

9. Shields, R.E., Korol, S., Carleton, R.N., et al. Brief mental 
health disorder screening questionnaires and use with 
public safety personnel: a review. Int J Environ Res Pub-
lic Health. 2021;18(7):3743. https://doi.org/10.3390/
ijerph18073743.

10. Singha, R., Singha, S. Mental health treatment: exploring 
the potential of augmented reality and virtual reality. In 
Applications of Virtual and Augmented Reality for Health 
and Wellbeing; Hiran, K.K., Doshi, R., Patel, M., eds. IGI 
Global: Pennsylvania, USA; 2024; pp.91–110. https://doi.
org/10.4018/979-8-3693-1123-3.ch006.

11. Bakır, Ç.N., Abbas, S.O., Sever, E., et al. Use of augmented 
reality in mental health-related conditions: a systematic 
review. Digit Health. 2023;9:20552076231203649. https://
doi.org/10.1177/20552076231203649.

12. Lecomte, T., Potvin, S., Corbière, M., et al. Mobile apps 
for mental health issues: meta-review of meta-analyses. 
JMIR Mhealth Uhealth. 2020;8(5):e17458. https://doi.
org/10.2196/17458.

13. Kernkraut, A.M., Diniz Nagem Janot de Matos, L. The impact of 
negative effects on the high prevalence of anxiety, depression 
and stress in healthcare professionals. Work. 2024; 79(2), 
857–866. Available Online: https://journals.sagepub.com/doi/
abs/10.3233/WOR-230257?download=true#:~:text=It%20
was%20observed%20that%20a,of%20depression%20
and%20stress%20prevalence.

https://doi.org/10.1080/17538157.2021.1983578
https://doi.org/10.1080/17538157.2021.1983578
https://doi.org/10.1093/jamia/ocae074
https://doi.org/10.1080/17434440.2021.2013200
https://openreview.net/pdf?id=d4UiXAHN2W
https://doi.org/10.1007/s11920-023-01438-4
https://doi.org/10.1007/s11920-023-01438-4
https://doi.org/10.3390/su15107905
 https://doi.org/10.1111/camh.12627
 https://doi.org/10.1111/camh.12627
https://doi.org/10.1016/j.jpsychores.2009.05.003
https://doi.org/10.1016/j.jpsychores.2009.05.003
https://doi.org/10.1007/s11126-017-9535-6
 https://doi.org/10.20944/preprints202307.2142.v1
 https://doi.org/10.20944/preprints202307.2142.v1
https://doi.org/10.18653/v1/2023.emnlp-main.370
https://doi.org/10.1093/jamia/ocae045
https://doi.org/10.1093/jamia/ocae045
https://doi.org/10.3390/ijerph18073743
https://doi.org/10.3390/ijerph18073743
https://doi.org/10.4018/979-8-3693-1123-3.ch006
https://doi.org/10.4018/979-8-3693-1123-3.ch006
https://doi.org/10.1177/20552076231203649
https://doi.org/10.1177/20552076231203649
https://doi.org/10.2196/17458
https://doi.org/10.2196/17458
https://journals.sagepub.com/doi/abs/10.3233/WOR-230257?download=true#:~:text=It%20was%20observed%20that%20a,of%20depression%20and%20stress%20prevalence
https://journals.sagepub.com/doi/abs/10.3233/WOR-230257?download=true#:~:text=It%20was%20observed%20that%20a,of%20depression%20and%20stress%20prevalence
https://journals.sagepub.com/doi/abs/10.3233/WOR-230257?download=true#:~:text=It%20was%20observed%20that%20a,of%20depression%20and%20stress%20prevalence
https://journals.sagepub.com/doi/abs/10.3233/WOR-230257?download=true#:~:text=It%20was%20observed%20that%20a,of%20depression%20and%20stress%20prevalence

