Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI
Generative AI for Transformative
Healthcare: A Comprehensive Study of
Emerging Models, Applications, Case
Studies and Limitations
SIVA SAI1, AANCHAL GAUR2, REVANT SAI3, VINAY CHAMOLA1(Senior Member, IEEE),
MOHSEN GUIZANI4(FELLOW, IEEE) and JOEL J. P. C. RODRIGUES5, (FELLOW, IEEE)
1Department of Electrical and Electronics Engineering, Birla Institute of Technology & Science (BITS), Pilani 333031, Rajasthan, India (e-mails:
{p20220063,vinay.chamola}@pilani.bits-pilani.ac.in)
2Department of Electrical and Communication Engineering, Maharaja Agrasen Institute of Technology, New Delhi, India (e-mail: anchal.avm@gmail.com)
3Department of Computer Science and Engineering, Birla Institute of Technology & Science (BITS), Pilani 333031, Rajasthan, India (e-mail:
f20212536@pilani.bits-pilani.ac.in )
4Department of Machine Learning , Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi 999041, UAE (e-mail: mguizani@ieee.org))
5COPELABS, Lusófona University, Lisbon, Portugal
This work was supported by CHANAKYA Fellowship Program of TIH Foundation for IoT & IoE (TIH-IoT) received by Dr. Vinay
Chamola under Project Grant File CFP/2022/027.
ABSTRACT Generative artificial intelligence (GAI) can be broadly described as an artificial intelligence
system capable of generating images, text, and other media types with human prompts. GAI models like
ChatGPT, DALL-E, and Bard have recently caught the attention of industry and academia equally. GAI
applications span various industries like art, gaming, fashion, and healthcare. In healthcare, GAI shows
promise in medical research, diagnosis, treatment, and patient care and is already making strides in real-
world deployments. There has yet to be any detailed study concerning the applications and scope of
GAI in healthcare. Addressing this research gap, we explore several applications, real-world scenarios,
and limitations of GAI in healthcare. We examine how GAI models like ChatGPT and DALL-E can be
leveraged to aid in the applications of medical imaging, drug discovery, personalized patient treatment,
medical simulation and training, clinical trial optimization, mental health support, healthcare operations
and research, medical chatbots, human movement simulation, and a few more applications. Along with
applications, we cover four real-world healthcare scenarios that employ GAI: visual snow syndrome
diagnosis, molecular drug optimization, medical education, and dentistry. We also provide an elaborate
discussion on seven healthcare-customized LLMs like Med-PaLM, BioGPT, DeepHealth, etc.,Since GAI is
still evolving, it poses challenges like the lack of professional expertise in decision making, risk of patient
data privacy, issues in integrating with existing healthcare systems, and the problem of data bias which are
elaborated on in this work along with several other challenges. We also put forward multiple directions for
future research in GAI for healthcare.
INDEX TERMS Generative AI, ChatGPT, Healthcare, LLMs, Applications.
I. INTRODUCTION
Tools based on artificial intelligence have gradually increased
in recent decades, and generative AI has emerged as a pow-
erful tool within this landscape. Generative AI combines ma-
chine learning techniques, deep neural networks, and natural
language processing (NLP) to learn patterns and character-
istics from vast datasets and generate outputs that resemble
human-generated content. The output can be generated in
various forms, such as audio, video, and text, depending upon
the demand. ChatGPT, developed by Open AI, is a language
model capable of generating human-like responses to text
inputs. It is built upon a transformer model and is one of the
most popular GAI models [1]. GAI models such as DALL-E
[2], Midjourney [3], and Stable diffusion [4] are capable of
generating high-quality images from textual prompts. These
GAI models also showcase the ability to bridge the gap
VOLUME X, 2020 1
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
between modalities [5] and aid education [6].
The GAI technology has shown promising potential in
healthcare. It can revolutionize how we approach medical
diagnosis, treatment, and patient care. The GAI models can
assist healthcare professionals in making clinical decisions
in various fields such as urology, radiology, and cardiology
[7]. A very recent survey by Market.us [8] predicts that GAI
in the healthcare industry is set to reach around $17 billion
by 2032, primarily driven by the automation in healthcare
operations of medical imaging and diagnostics and drug
discovery and development.
To generate reliable results in healthcare, GAI models need
to train on a large volume of medical data, including patient
records, medical images, and genomic sequences. These
trained models can provide innovative solutions to traditional
problem-solving and augment healthcare professionals’ ca-
pabilities to enhance patient outcomes. GAI models can also
simulate and predict disease progression, thus further helping
in better understanding and monitoring. A study by Messiah
et al. indicates the reliability of such models. The GAI model
was able to answer all queries regarding typical clinical tox-
icology cases of acute organophosphate poisoning [9]. The
GAI technology can also be used in disease management and
risk assessment, as well as to increase research education and
drug development. It has opened the windows to innovative
healthcare using technology [10].
One of the major areas where GAI is making a significant
contribution to healthcare is medical imaging. GAI models
like DALL-E can assist in detecting and diagnosing diseases
by analyzing patient medical images, such as X-rays, MRIs,
and CT scans. GAI algorithms can be trained to learn and
identify subtle patterns and anomalies in scans that often slide
past the naked eye. By using generative models, timely inter-
ventions and improved patient care can be provided. With
higher accuracy and speed, this technology helps in early
disease detection, such as cancer or neurodegenerative dis-
orders. Furthermore, GAI can augment healthcare research
and education. The models help in experimentation and hy-
pothesis by generating synthesized data. They can produce
virtual patient scenarios, enabling more practical education
for healthcare professionals. GAI models can help Biomarker
identification by analyzing large-scale genomic, proteomic,
or imaging data. They facilitate research and study by gen-
erating synthetic data on which medical researchers can
perform experimentation. GAI models can generate synthetic
data samples that exhibit specific biomarker characteristics
to help study pathology, which helps researchers visualize
complex medical data and facilitate exploratory analysis for
better understanding.
In addition to medical imaging and research, GAI has vari-
ous other applications, including drug development, chatbots,
personalized patient treatment, medical simulation and train-
ing, clinical trial optimization, and mental health support.
Healthcare professionals can use various GAI models like
ChatGPT for assistance in diagnosis and treatment. It is seen
that GAI models like ChatGPT can learn and identify their
own mistakes just by prompting it to check if any output is
wrong. These models can also generate patients’ discharge
summaries by leveraging a large amount of data they are
trained on and by analyzing patient data and their medical
records. They can do so without any detailed description or
meaning of medical terms provided beforehand. This paper
discusses how these and other generative models can enhance
the healthcare system.
The integration of GAI and healthcare also presents mul-
tiple challenges. Data privacy, ethical considerations, and
data bias are critical aspects that need attention in this GAI-
healthcare confluence to utilize the technology’s full poten-
tial while ensuring patient safety and security. Therefore,
in this work, the contributions are summarized as follows:
1. Discuss how GAI can support healthcare systems, high-
lighting their limitations and how to overcome them. 2.
Analyze some real-world GAI models in healthcare systems.
3. Provide a variety of applications of GAI in healthcare. 4.
Describe four real-world scenarios of using generative AI
in healthcare. 5. Discuss seven healthcare-customized GAI
models. 6. Present several limitations and future directions
on the applications of GAI in healthcare.
A. ORGANIZATION
The rest of the paper is organized as follows. Section II pro-
vides a brief overview of generative AI. Section III presents
an analysis of the real-world performance of GAI models
in healthcare. In Section IV, we provide and elaborate on
a variety of applications of GAI in healthcare.Section VI
describes four real-world scenarios of using generative AI in
healthcare - visual snow syndrome, molecular optimization,
medical education, and dentistry. Section V discusses seven
healthcare-customized GAI models. Sections VII and VIII
present limitations and future research directions on GAI
applications in healthcare, respectively. Finally, the review is
concluded in Section IX.
II. AN OVERVIEW OF GAI
Generative AI is an advanced artificial intelligence tech-
nology that has recently gaining significant attention and
corporate funding. Its popularity has led to different startups
being formed solely on the development of GAI technology
[11]. In response to user prompts, GAI models can generate
various forms of media, including text, images, audio, video,
and 3D models. This cutting-edge technology harnesses the
power of pattern recognition and learns from existing data to
generate novel and distinctive results that closely resemble
the characteristics of the input training data. GAI has rapidly
gained popularity and is now widely regarded as one of the
most coveted technologies in the world.
What sets GAI apart is its ability to produce realistic
and coherent outputs. Unlike traditional AI systems designed
for specific tasks, GAI surpasses rule-based and determin-
istic approaches. It extensively utilizes advanced machine
learning techniques such as Deep Learning (DL), Natural
Language Processing (NLP), and Neural Networks. These
2VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
techniques enable systems to discern patterns and traits from
vast training datasets, empowering them to generate new
data that closely resembles the original information. GAI
models are unique, showing enhanced creativity and novelty
in generating data. The data produced is not just a copy of
training data but something different with its original traits.
They can train on unlabeled data and map underlying patterns
and structures independently. This ability of unsupervised
learning makes GAI models valuable when labelled data is
scarce.
GAI models have seen a sharp rise in usage and pro-
duction. Over time, GAI models have become more so-
phisticated, employing complex architectures with improved
stability and quality in generating realistic data in different
modalities. Techniques like conditional generation in GANs
and fine-tuning language models enable more precise and
controllable content generation.
Noteworthy examples of GAI systems include ChatGPT,
Dall-E, Midjourney and Bard. ChatGPT, developed by Ope-
nAI, is one of the most popular GAI models known for its
natural language processing capabilities. It engages users in
coherent and contextually relevant conversations.
Large Language Models (LLMs) represent a transforma-
tive breakthrough in natural language processing (NLP),
marked by their immense scale, complex architecture, and
remarkable language generation capabilities. ChatGPT is
a generative pre-trained transformer and belongs to the
family of large language models. GPT utilizes a decoder-
only transformer architecture. This architecture enables it
to probabilistically generate sequences of words or tokens,
given an input prompt or context. The model predicts the
most likely sequence of words following the input based
on the patterns learned during training. It relies heavily
on self-attention mechanisms. The transformer architecture
facilitates the model’s ability to process sequential data ef-
ficiently by simultaneously attending to different parts of
the input sequence. This mechanism allows the model to
capture dependencies and relationships between words in
long-range contexts, which is crucial for understanding and
generating coherent text. Notably, the technical prowess and
generative prowess of GPT have set new benchmarks in the
field, showcasing its adaptability and performance without
extensive fine-tuning for specific tasks. GPT models, such as
GPT-3 [12], have enormous parameters, often in the billions,
allowing them to capture intricate linguistic nuances and
context. This large parameter count contributes to their abil-
ity to generate diverse and contextually relevant text across
various domains. However, LLMs’ scale and computational
demands, like GPT, pose challenges regarding resource re-
quirements and potential biases inherited from the training
data. Further research is ongoing to optimize these models
for efficiency and mitigate biases. Nevertheless, the advent
of Large Language Models, particularly exemplified by GPT,
stands as a monumental advancement in machine learning,
revolutionizing the capabilities of NLP systems and paving
the way for increasingly sophisticated language understand-
ing and generation technologies.
Figure 1 displays existing GAI models. As the figure
shows, GAI models can generate various data types such as
audio, video, text, images, 3D visual and code. These GAI
models can generate intricate content that mirrors human
creativity. This characteristic makes GAI an invaluable tool
which can be used in various industries, including gaming,
entertainment, and product design. Over the past decade,
this has led many multinational corporations like Google,
Microsoft and numerous smaller firms to invest in actively
developing and refining this technology. Dall-E is another
GAI model developed by OpenAI. It produces images based
on textual prompts. Its potential can be extended into health-
care in many ways. It can be used in medical imaging to assist
radiologists and clinical workers, as this model can be trained
on different medical image data and their textual descriptions
and generate relevant synthetic data. It is important to note
that GAI is a rapidly evolving field with ongoing research
and experimentation to develop this technology further.
III. REAL WORLD GAI PERFORMANCE IN HEALTHCARE
GAI has shown great development in recent years, demon-
strating remarkable capabilities in various applications, from
text to images. However, while GAI has shown great promise
in controlled environments, assessing the GAI models in
real-world scenarios is essential to test the model’s relia-
bility and effectiveness. While evaluating the performance
of GAI models outside a controlled model, several factors
need to be considered, including the reliability of outputs,
bias and fairness, generalization across different populations,
interpretability, and the potential impact on human decision-
making. Testing requires rigorous evaluation methodologies
and comprehensive datasets.
An evaluation was done on ChatGPT, a large language
model (LLM) by Kung et al. [13].They tested ChatGPT to
answer the United States Medical Licensing Exam (USMLE)
questions. The test was conducted in three stages of stan-
dardized tests, and ChatGPT could pass all the stages without
specific training or reinforcement. It is important to note that
these tests were of expert-level knowledge, and ChatGPT was
able to pass with 60% accuracy. Furthermore, the answers
provided by ChatGPT did not require any specialized input
and were detailed while including clinical insights and com-
prehensive reasoning.
Another evaluation was done on ChatGPT 4 [14], an
upgraded version of ChatGPT by Teebagy et al. [15]. They
tested the model on the Ophthalmology Knowledge As-
sessment Program (OKAP) examination conducted by the
American Academy of Ophthalmology (AAO) to assess res-
idents’ knowledge of ophthalmology training programs. The
examination consisted of 180 questions and content of ques-
tions ranging from anatomy and physiology to ophthalmic
subspecialties such as cornea and neuro-ophthalmology [16].
Results indicated that ChatGPT 4 could pass the examination
and outperform its predecessor, ChatGPT 3.5, thus establish-
ing the growing potential for using GAI models in healthcare
VOLUME X, 2020 3
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
TEXT AUDIO VIDEO IMAGE 3D VISUAL CODE
GENERATIVE AI MODELS
Murf AI Voice
Generator
ChatGPT Google
Imagen Video
OpenAI-Dall-
E2
OpenAI-Point-
E
DeepMind
AlphaCode
Google's
BERT
Stability Stable
Diffusion
OpenAI
MuseNet
Krikey AI
platform
OpenAI ChatGPT
& ChatGPT 4
Google Dream
Fusion
CTRL Soundful AI
Music Generator
Meta Make a
Video Midjourney Microsoft
Rodin Diffusion
FIGURE 1: Popular GAI models
consultation and treatment. ChatGPT also passed the radiol-
ogy board–style examination nearly [17]. The examination
of 150 questions without images of multiple choice answers
was conducted with the questions of the difficulty level of
Canadian Royal College and American Board of Radiology
examinations. ChatGPT gained an overall score of 69% by
correctly answering 104 questions out of 150. It showed
better performance in clinical management questions and
low-order thinking questions. Table 1 shows that GPT 4 can
handle multi-level prompts, perform complex analysis and
provide more detailed results compared to its predecessor
ChatGPT-3 [18].
The performance evaluation of ChatGPT showed the po-
tential reliable use of GAI in healthcare. With the GAI
models currently evolving and under rigorous research and
development, it opens a wide window for integrating various
GAI models in the daily dealings of healthcare professionals.
The GAI technology will soon be positioned to be a part
of regular clinical practice, contributing in various lengths,
accounting for its wide applications in various healthcare
fields, and enhancing patient care.
IV. APPLICATIONS
This section discusses the applications of using GAI in
different healthcare spheres. Figure 2 presents an overview
of the discussed applications.
A. MEDICAL IMAGING
Medical imaging is a rich and non-invasive technique that
provides healthcare professionals with a detailed visualiza-
tion of the patient’s anatomical structures for treating medical
conditions. This technique enables early detection of dis-
eases, improves screening procedures, and guides treatment
planning strategies. In the current scenario, medical imag-
ing faces many limitations, such as insufficient annotated
data and limited imaging modalities and contrast. GAI has
emerged as a promising solution to address these challenges
and further enhance medical imaging capabilities.
1) Data augmentation
It is often challenging to train deep learning models as the
datasets available in medical imaging are limited. GAI tech-
niques, such as Generative Adversarial Networks (GANs)
and Variational Autoencoders (VAEs), provide a reliable
solution as they can generate synthetic images that resemble
patient data, increasing the models’ robustness. GAI models
are also capable of generating data on rare conditions such as
4VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Applications of Generative AI for
healthcare
Personalised
Patient Treatment
Medical
Imaging
Text
Generation and
Summarization
Drug
Discovery and
Development
Clinical Trial
Optimization
Mental
Health Support
Chatbots Medical
Simulation and
Training
Human
Movement
Simulation and
Analysis
Healthcare
Operations and
Resource
Management
FIGURE 2: Applications of Generative AI in Healthcare
Aquagenic Urticaria [19], methemoglobinemia [20] and gen-
erate missing data, thus providing an augmented and diverse
dataset for training and evaluation, enhancing the ability to
detect and diagnose diseases in real-world scenarios. It is
observed that while GAI methods like GANs and VAEs play
a pivotal role in addressing limited annotated data in medical
imaging, the extent of their utilization might vary due to the
prevalence of scarce datasets for rare conditions.
2) Image enhancement and reconstruction
One of the essential elements in disease diagnosis through
medical imaging is the image quality of scans such as an X-
ray or CT scan. When this image suffers distortion due to
noise, missing data, and low resolution, it can lead to misin-
terpretations and delays in the diagnosis. GAI models can be
trained to remove noise from these images using frameworks
such as generative adversarial networks and autoencoders;
this improves the accuracy of quantitative analysis, allow-
ing reliable measurements and quantitative parameters that
eventually aid in assessing disease progression, longitudinal
monitoring and treatment response. Diffusion-based model
DiffMIC [21] is tailored for medical image classification. It
focuses on eliminating noise and perturbations while robustly
capturing semantic representations. DiffMIC uses a dual con-
ditional guidance strategy that enhances step-wise regional
attention by conditioning each diffusion step with multiple
granularities. Additionally, this study [21] proposes a method
to learn mutual information within each granularity by en-
forcing Maximum-Mean Discrepancy regularization during
the diffusion forward process. GAI models can also produce
high-resolution images from low-resolution data by applying
a super-resolution approach. The super-resolved images are
used for accurate disease diagnosis and detailed study of
the anatomical structures. Bing et al. [22] used an enhanced
generative adversarial network for the super-resolution re-
construction of images. The authors improved the squeeze
and excitation blocks in GANs generator and discriminator
by strengthening important features and weakening the non-
important ones. Furthermore, they used low function loss to
train the model by combining L1 loss, mean square error loss,
perceptual loss, and relativistic adversarial loss.
Furthermore, challenges like varying sequence lengths,
missing data or frames, and high dimensionality pose sig-
nificant hurdles for conventional models. A novel approach
called Sequence-Aware Diffusion Model (SADM) [23] is
introduced for generating longitudinal medical images to
address this challenge of modeling the dynamic anatomy
of the human body, which can be influenced by both long-
term (e.g. chronic diseases) and short-term (e.g. heartbeat).
It introduces a sequence-aware transformer as the condi-
tional module within the diffusion model. This innovative de-
sign enables learning longitudinal dependencies, even amidst
missing data during training. Furthermore, it enables the
autoregressive generation of image sequences during infer-
ence, offering a more comprehensive insight into anatomical
changes over time.
3) Cross-Modality Image Translation
There are many ways to perform medical imaging. These
methods include techniques such as X-ray, positron emis-
sion tomography (PET), magnetic resonance imaging (MRI),
computed tomography (CT), ultrasound, and Single-Photon
VOLUME X, 2020 5
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Emission Computed Tomography (SPECT), to name a
few. These different techniques are called modalities. Each
modality uses different physical properties and imaging tech-
niques to generate images; each has different advantages
and limitations regarding image quality, spatial resolution,
sensitivity, and contrast.
Cross-modality image translation converts medical images
from one modality to another while preserving relevant fea-
tures. This enables the fusion of modality-specific advantages
and complementary information; it addresses data limitations
and offers clinical flexibility. This flexibility allows several
diagnostic choices, but it is a challenge to translate infor-
mation from one modality to another. GAI techniques, such
as generative adversarial networks (GANs) and variational
autoencoders (VAEs), have shown remarkable capabilities in
cross-modality image translation. They excel at learning the
complex mapping between different image modalities. They
preserve underlying structural and pathological information,
facilitating accurate cross-modality translation. Generative
AI also helps overcome modality-specific limitations, such
as if a modality has spatial resolution or contrast issues.
GAI can transform images from higher resolution or con-
trast modality to enhance visualization. It helps healthcare
professionals make more clinically informed decisions; they
can leverage the benefits of multiple modalities to produce
advanced imaging to facilitate diagnosis and treatment.
Generative AI-based cross-modality translation opens up
new avenues for research and innovation in medical imag-
ing. It has proved useful in generating synthetic images,
preserving important information, promoting research and
innovation, and fusing inter-modality data. The GAI has
unlocked the full potential of diverse image modalities by
which healthcare professionals can make better diagnoses
and devise personalized patient treatment strategies.
4) Interpretability and Explainability
GAI models help achieve interpretability and explainabil-
ity in medical imaging [24]. Deep learning models often
lack this, making understanding the reasoning behind their
predictions challenging. Interpretability refers to the ability
to understand and explain the decision-making process of
the AI model. In medical imaging, interpretability is crucial
because healthcare providers and patients must trust and un-
derstand the generated results to proceed with treatment and
make informed decisions [25]. Various methods can achieve
this [26], the most important being visualization. Learning
features are visualized to understand which image regions
or characteristics influence the generated output. For this,
attention mechanisms are also used, explicitly highlighting
the important regions or features that contributed to the
output; this provides transparency and a clear understanding
to healthcare professionals as to why certain areas are em-
phasized. Explainability tells the user why a specific decision
was made by providing a clear and coherent explanation.
It makes the decision-making process more transparent and
easy to understand for the users, allowing them to put better
trust in the model. Some explainability techniques help the
model achieve this. Local explanation focuses on explain-
ing the decision on a specific image instance; this helps
understand the model’s decision at every level. Keeping the
model architecture and decision-making process simple and
transparent can enhance explainability. GAI aids in achiev-
ing interpretability and explainability, which are crucial for
trust and understanding in medical AI. Techniques such as
attention mechanisms and local explanations contribute to
transparency; however, ensuring a balance between model
complexity and explainability remains challenging.
B. DRUG DISCOVERY AND DEVELOPMENT
Drug development is associated with bringing new therapies
to the market. It is a complex and time-consuming pro-
cess involving high costs and low success rates. GAI offers
a promising solution to the de novo design of molecular
structures. It can generate novel compounds, optimize drug
candidates and predict the properties of drugs. GAI is capable
of 1D, 2D and now 3D models of molecules [27]. Inception
Score (IS) was proposed by Salimans et al. for generative
models [28], which investigates if the generated molecules
can be classified correctly to cover the chemical space de-
fined by the training set.
Among various challenges in drug development is the
prediction of protein function, an area that has seen sig-
nificant advancement through various machine-learning ap-
proaches in recent years. However, prevalent methods often
frame this task as a multi-classification problem, assigning
predetermined labels to proteins. Prot2Text [29] is a novel
approach that departs from the traditional binary or cate-
gorical classifications by predicting protein functions in a
free-text format. This innovative methodology employs an
encoder-decoder framework that integrates Graph Neural
Networks (GNNs) and Large Language Models (LLMs).
Through this amalgamation, diverse data types, including
protein sequences, structures, and textual annotations, are
effectively assimilated, ensuring a comprehensive represen-
tation of proteins’ functions and generating detailed and
accurate descriptions. To evaluate the efficacy of Prot2Text,
a multimodal protein dataset was curated from SwissProt,
demonstrating its effectiveness through empirical analysis.
The results underscore the transformative potential of mul-
timodal models, particularly the fusion of GNNs and LLMs,
which equips researchers with potent tools for precise predic-
tions of proteins’ functions.
The GAI also assists in understanding the structure-
activity relationship (SAR) of the molecules. It generates a
diverse set of molecules and studies how the activity changes
corresponding to their structure which helps researchers get
insights into chemical features essential for drug-target inter-
actions.
The GAI also facilitates generating new molecules with
different core structures while retaining key pharmacophoric
features known as scaffold hopping. It enables the re-
searchers to explore the chemical space beyond existing
6VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Prompts ChatGPT GPT 4
Prompt 1: Diagnosing a patient
showing ambiguous symptoms
A Patient presents with occasional dizziness,
weight loss, fatigue and low blood pressure.
What are some possible causes of these
symptoms?
A 30 year old female with a 2 month
history of unintentional weight loss
of about 10 pounds, progressive fatigue
and episodes of dizziness. Please provide
differential diagnosis and suggest relevant
diagnostic tests.
Prompt 2: Patient education Explain Tuberculosis in simple terms.
Create a patient friendly handout on
tuberculosis, including an overview of
the condition, symptoms, risk factors,
potential complications, and management
strategies.
Prompt 3 – Reviewing medical
research
Tell me about the benefits of exercise in
improving mental health.
Define the relationship between physical
exercises and mental health by summarizing
recent research findings. Include the
influence of different types of exercise and
recommendations for various populations.
TABLE 1: Performance of GPT 4 over ChatGPT based on user prompts
scaffolds, potentially leading to improved drug candidates
with different properties and mechanisms of action.
C. PERSONALIZED PATIENT TREATMENT
GAI has advanced significantly in personalized patient treat-
ment, evolving from initial predictive modeling to integrat-
ing specific conditions for tailored treatment plans. Condi-
tional Variational Autoencoders (CVAEs) are GAI models
that combine conditional variables and variational autoen-
coders to learn latent patient information while incorporating
specific conditions relevant to personalized treatment. By
altering values of different conditions, the model generates
various treatment plans tailored to individual patients [30].
Fine-tuned NLP models such as Bidirectional Encoder Rep-
resentation from Transformers (BERT) have improved in
generating personalized treatment summaries and adaptive
plans based on patient data.
D. MEDICAL SIMULATION AND TRAINING
Medical trainees and professionals must refine their clinical
skills in a controlled environment; this is done via medical
simulation. Realistic and immersive experiences of real-life
critical conditions can be simulated using GAI. GAI tech-
niques such as virtual patient simulation, procedural simula-
tions, scenario generation, and haptic feedback can transform
medical education.
GAI models can be employed to generate virtual avatars
of patients. These avatars resemble closely to the patients
and can be customized by adjusting parameters such as age,
gender, and medical history, allowing for a more realistic
experience. StyleGAN2 (Style-based Generative Adversarial
Network) model can generate high-resolution images with
realistic details; these models can be adapted to generate
virtual avatars of the patient [31]. By incorporating data from
physiological models and clinical knowledge, the GAI can
create fundamental changes in organ function, vital signs
and disease progression over time by capturing the tem-
poral dynamics and complications associated with specific
conditions. It not only helps medical professionals simulate
complicated surgery but also helps train medical students.
The generative models can expose healthcare professionals
to a wide range of patient cases, sudden deterioration, and
adverse reactions. The different scenarios are designed to
train them to handle complex and unpredictable situations,
enhancing their decision-making skills and clinical compe-
tence. The generative models can generate rare and uncom-
mon conditions; this allows students to Gain exposure and
develop proficiency in dealing with such cases.
GAI models can generate realistic haptic feedback by
analyzing visual and contextual information. The haptic sig-
nals mimic the tactile sensations experienced during medical
procedures. The generative models can also generate 3D
representations of organs, bones, or blood vessels, which
can be integrated into haptic feedback systems allowing
professionals to interact with and manipulate them as if they
were real.
E. CLINICAL TRIAL OPTIMIZATION
Introducing and evaluating new interventions and translating
research findings into clinical practice is essential. They al-
low head-to-head comparisons of different treatment options,
providing researchers with valuable evidence for choosing
the most appropriate treatment for specific patient popula-
tions. However, the complexity and challenges of recruiting
a diverse patient pool while ensuring safety make clinical
trials tedious. GAI presents a transformative approach to
address these complexities and optimize clinical trials. GAI
helps advance protocol design and validation for clinical
trials [32]. It does so by simulating virtual trials with differ-
ent designs, randomizing strategies or building an inclusion
VOLUME X, 2020 7
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
criterion. GAI helps researchers reduce biases, predict treat-
ment responses based on patient characteristics, and simulate
different scenarios. Generative models like DeepSurv and
DeepHit can predict patient responses based on different
patient characteristics and genetic information; these models
can evaluate the potential outcomes of different interventions,
which aids professionals in understanding how different
subgroups react to specific conditions. GAI helps estimate
sample size for clinical trials by running multiple iterations
of simulated virtual clinical trials using synthetic patient
populations. Dreesbach et al. [33] proposed a new approach
for clinical trials using longitudinal clinical study data by
employing the Variational Autoencoder Modular Bayesian
Network (VAMBN) model. Virtual patient data was gener-
ated while making theoretical guarantees on data privacy. It
could help in trial design and facilitate data sharing. The GAI
can successfully select and optimize endpoints for clinical
trials. It can identify clinical outcomes and endpoints by an-
alyzing historical data and meaningful patterns for patients,
researchers and regulatory agencies. The GAI in clinical trial
optimization can significantly enhance trial efficiency, im-
prove patient stratification, reduce costs and generate reliable
and generalizable evidence. Researchers can optimize trial
protocols using the GAI to personalize treatment and improve
patient care.
F. MENTAL HEALTH SUPPORT
Mental health is of utmost importance in leading a healthy
and peaceful life. Any disruption in mental health directly
impacts the overall well-being of an individual. Today, prob-
lems like anxiety, frustration, and depression are common
worldwide. The GAI is a tool that can help the stressed
population to improve their quality of life. The GAI gives
personal treatment plans and therapy based on individual
needs. They can perform sentimental analysis by analyzing
text and speech patterns and detecting sentimental and emo-
tional cues. The GAI contributes to the early detection of
mental health conditions by analyzing large volumes of data
and identifying patterns and indicators suggestive of specific
mental health conditions and can flag individuals who may
be at a higher risk or may require immediate assistance.
Yang et al. [34] use generative adversarial networks and
hierarchical attention mechanisms to diagnose depression us-
ing multi-modal data, including text and audio physiological
signals. Generative AI can generate an immersive virtual
reality environment for mental health patients, providing a
safe time off from the real world [35]. A correct dosage of
this can significantly improve their conditions; the virtual
reality environment can simulate exposure therapy, relaxation
techniques and stress management scenarios. By leveraging
GAI capabilities, mental health support can be more acces-
sible, effective and personalized. With human oversight, the
GAI can augment existing mental health services. Therefore,
Leveraging GAI capabilities augments existing mental health
services, making them more accessible, effective, and per-
sonalized. With human oversight, GAI tools can supplement
and enhance mental health support by providing tailored in-
terventions, ultimately aiding in improving the overall well-
being of individuals. [36]
G. HEALTHCARE OPERATIONS AND RESOURCE
MANAGEMENT
Healthcare operations and resource management are complex
and multifaceted areas which involve various aspects such
as resource allocation, demand prediction, workflow opti-
mization, and much more. Largely, these tasks are driven by
human decision-making, which can be time-consuming and
prone to errors; given the medical aspect, it can also be fatal
in some instances. However, with the advent of GAI models,
healthcare organizations can harness GAI’s power to make
data-driven decisions and enhance operational efficiency. A
generative model, DALL-E, can be used to generate visual
representations of healthcare facility layouts and floor plans
to help identify areas of improvement, streamline workflows
and ensure optimum utilization of space and equipment. It
can improve communication and understanding by generat-
ing visual aids for patients with specific needs or language
barriers. It also generates illustrations for standard operating
procedures (SOPs) or guidelines on different healthcare con-
ditions to train staff and ensure consistent practices. Another
generative AI model, ChatGPT, can assist in operations man-
agement by scheduling appointments with patients, integrat-
ing with hospital information systems, and providing real-
time updates on wait times. GAI models can automate routine
decision-making processes such as approving and denying
and even recommending [37] different procedures based
on predefined criteria or conditions, reducing administrative
burden and enabling a faster response time.
H. MEDICAL CHATBOTS
One of the significant applications of GAI has been the
development of medical chatbots. Chatbots are versatile com-
puter programs designed to simulate human conversation
and automated responses. They can serve as virtual assis-
tants for patient support and engagement. They can answer
common questions, offer guidance and provide medication
reminders. ChatGPT, based on the GPT (Generative Pre-
trained Transformer) architecture, can be trained on a wide
range of healthcare data, enabling it to provide accurate and
consistent responses to patient queries. Dave et al.. [38] dis-
cuss various applications of ChatGPT in medicine, including
its aid to patients as a virtual assistant and record keeping
of patient files. Patients usually seek information about any
symptom they are showing or a specific medical condition;
ChatGPT can offer preliminary guidance by asking relevant
questions. Based on the information provided by the patient,
ChatGPT can assess the urgency and severity of the symptom
and provide general information about potential causes and
common conditions associated with those symptoms, self-
care measures and when to seek medical attention; in this
way, they assist in triage and symptom assessment.
8VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Chatbots can be a reliable source of health information and
education; they can explain medical terms Moreover, it offers
guidance on various health topics. Information provided by
ChatGPT is reliable as it passed the United States USMLE,
[13] which is a medical licensing exam; the exam was con-
ducted in three levels, in which ChatGPT could perform near
the passing threshold and gave answers with concordance and
insights in its explanations. Lee [39] explores the potential of
ChatGPT in medical education as a virtual assistant to teach
medical students and increase their engagement and learning.
Recently, DiagnaMed Holdings Corp. released Dr GAI, a
GAI medical chatbot based on ChatGPT, which helps people
in a general health advisory. Dr GAI is the third commercial
product from the company’s Health GAI division, which
focuses on building generative AI applications for healthcare.
Another chatbot, InstructGPT [40], a ChatGPT variant,
can provide step-by-step instructions to patients. It is advan-
tageous in providing medication administration instructions
to patients. It can generate step-by-step guides on admin-
istering different medications, including information from
opening packages to measuring dosages and using specific
medical devices such as syringes or droppers. InstructGPT
can also generate medical schedules for patients depend-
ing on their prescriptions. However, it is essential to note
that chatbots like ChatGPT and InstructGPT should only
work as a tool for assistance under clinical supervision.
Seeking proper medical guidance from certified healthcare
professionals is essential in all medical conditions. Hence,
GAI-driven medical chatbots emphasize their versatility as
virtual assistants, aiding in triage, providing reliable health
information, supporting medical education, guiding patients
in medication administration, and highlighting the necessity
of clinical supervision in all medical conditions. At the
same time, while GAI-powered chatbots are beneficial as
assistance tools, they should always operate under clinical
supervision.
I. HUMAN MOVEMENT SIMULATION AND ANALYSIS
Understanding human movements is vital to get a detailed
picture of the dynamic anatomy of the human body. Human
movement simulation and analysis can be performed using
GAI technologies and get valuable insights into diagnostic,
therapeutic, and performance-based optimization processes.
Healthcare professionals can significantly improve treatment
planning and patient care by carefully analyzing human body
movements. Generative models utilize advanced machine-
learning techniques to simulate human movement with re-
markable accuracy and detail.
Midjourney can simulate movements tailored to individual
patients’ needs, helping in virtual rehabilitation and phys-
ical therapy sessions. It can provide interactive guidance
and visual demonstrations of correct movement techniques.
This allows better patient engagement and rehabilitation out-
comes. GAI models also help assess gait abnormalities, an-
alyze movement patterns and guide rehabilitation strategies
for individuals with mobility impairments, enabling tailored
interventions and objective progress monitoring.
By leveraging Midjourney and similar GAI models in
human movement simulation and analysis, healthcare profes-
sionals can access a powerful tool that accurately simulates
realistic movements, enhancing their understanding of move-
ment patterns and contributing to better patient care.
J. INSURANCE PRE-AUTHORIZATION/PRIOR
AUTHORIZATION
Pre-authorization, also known as prior authorization or pre-
approval, is a mechanism that healthcare payers (such as
insurance companies) use to decide whether to cover and
reimburse a certain medical operation, prescription, or ser-
vice. It entails getting the payer’s consent before a health-
care service is rendered or a course of treatment is begun.
Pre-authorization ensures that planned healthcare services
or treatments satisfy the payer’s requirements for coverage,
which may include medical necessity, appropriateness, cost-
effectiveness, and adherence to set standards or rules. It gives
payers a mechanism to keep expenses under control, limit
use, and guarantee that the services being rendered comply
with the conditions of the insurance or healthcare plan.
Prior authorization is one healthcare procedure that GAI
could potentially enhance. Even though the healthcare sector
has made progress towards automating and standardizing PA,
the procedure still presents administrative challenges. Re-
viewing PA requests requires a significant amount of clinical
staff time from payers. According to reports, doctors and staff
spend up to 13 hours each week on the PA process [41]; many
clinicians think this compromises their clinical judgement
and can delay providing timely care.
An in-depth evaluation by McKinsey and Company in-
dicates that GAI-enabled Pre-Authorization can automate
50 to 75 per cent [41] of manual activities, increasing ef-
ficiency, lowering costs, and enabling clinicians at payers
and providers to concentrate on challenging cases and actual
care delivery and coordination. As a result, both clinicians
and health insurance subscribers may have a better overall
healthcare experience.
The current pre-authorization workflow involves much
manual labour. Some payers have started the PA automation
path to increase productivity, decrease provider distrust and
unhappiness, and enhance doctor and client experiences. For
instance, electronic prior authorization quickens the data
flow between payers and providers. Electronic PA digitalizes
workflows, which speeds up turnaround times. With elec-
tronic PA, close to 60 per cent of requests were fulfilled
within two hours as opposed to zero requests sent by phone
or fax.
The GAI may help organize data from electronic health
records, emails, policies, medical procedures, and other
sources by utilizing advanced Machine Learning algorithms
like Natural Language Processing and digital and workflow
management technologies. This could significantly reduce
low-value, time-consuming tasks involving searching, com-
piling, and validating details that people previously per-
VOLUME X, 2020 9
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
formed manually. While GAI will undoubtedly change the
prior authorization process, several obstacles must first be
solved. Payers will first want unrestricted access to EHRs,
which requires rigorous adherence to data privacy laws and
a significant amount of design work to assure interoperabil-
ity among diverse EHR application software and platforms.
Another crucial requirement for AI-driven PA automation
is the definition of common criteria for attachments, data
blueprints, and information-sharing protocols by industry
participants at the operational level.
Although highly skilled physicians will always be the
ones to make the final pre-authorization decisions, GAI can
help payers and providers make critical decisions while also
increasing efficiency and improving provider and patient ex-
perience. Insurance companies can delegate the most difficult
and delicate decision-making to highly skilled doctors by
automating most pre-authorization decisions. For optimal use
of these advantages, stakeholders will need to collaborate to
design a new set of standards for data sharing and additional
protocols for system integration and interoperability.
K. MEDICAL TRIAGE
In medicine, triage is defined as prioritizing the care of
patients (or catastrophe victims) based on their condition,
severity, prognosis, and resource availability. Triage is used
to identify patients requiring rapid resuscitation, allocate
them to a designated patient care area to prioritize their care,
and start necessary diagnostic and therapeutic procedures
[42].
By offering helpful assistance and enhancing decision-
making, generative AI has the potential to impact the field
of medical triage significantly [43].Creating triage support
systems is an important example of generative AI’s use in
medical triage. These systems analyze patient data, includ-
ing symptoms, medical history, and test results, using deep
learning and pattern recognition to provide predictions and
suggestions on the criticality of therapy [44]. Generative AI
can help healthcare personnel make better-educated triage
decisions by quickly analyzing massive quantities of data.
Rapid triage can be made possible using the GAI in
emergencies where time is of the essence. Disaster events and
emergency departments sometimes include difficult circum-
stances and scarce resources. Healthcare practitioners can
input patient information into the system using GAI models,
which can quickly analyze the data and produce assessments
that help prioritize patients based on the extent of their condi-
tions. Assuring prompt care for individuals most in need can
greatly improve productivity, the distribution of resources,
and, ultimately, patient outcomes. Another field where GAI
can help is in addressing inequalities in triage judgements.
Implicit biases can have an impact on triage decisions, among
other things. GAI can reduce these discrepancies by offering
a data-driven, objective approach to triage. It is feasible to
lessen prejudice and increase fairness in the triage process
by training AI models on various representative datasets,
ensuring that patients receive the proper degree of treatment
based on their medical requirements rather than other criteria.
However, ethical issues are raised by using GAI in medical
triage. To ensure that triage decisions can be understood,
supported, and accepted by healthcare professionals and pa-
tients alike, transparency and explainability of AI models are
essential. Furthermore, appropriate security measures must
be implemented to defend patient privacy and guarantee the
secure handling of sensitive medical data. It is imperative
to understand that GAI should not take the role of human
expertise. Cooperation between artificial intelligence systems
and healthcare personnel is essential to properly integrate
GAI into the triage process and take advantage of both
parties’ strengths Theriseo67:online. Recent developments
in medical triage highlight the use of advanced AI, such as
GAI tools, aiming to expedite decision-making by analyzing
a wide range of patient information swiftly and accurately
[45]. Furthermore, Predictive analytics models incorporating
AI and machine learning are integrated into triage systems
to forecast patient outcomes, predict disease progression,
and identify high-risk patients. Medical triage can be greatly
enhanced by combining the strength of generative AI with
human judgement, improving patient care and utilization of
resources in challenging healthcare scenarios.
L. TEXT GENERATION AND SUMMARIZATION
Medical data are abundant in medical records, scientific
literature and patient feedback; extracting valuable informa-
tion is time-consuming if done manually. GAI models like
BART (Bidirectional and AutoRegressive Transformers) can
provide significant support. BART is a transformer architec-
ture that utilizes self-attention mechanisms to process and
understand text data. By analyzing patient data and context,
GAI models like BART can assist in generating clinical
documentation such as discharge summaries, progress notes
and medical reports that capture relevant information and
generate personalized recommendations for diagnosis and
treatment plans. BART reduces the burden on healthcare
providers, allowing them to focus more on patient care; it
can help generate patient-friendly explanations of medical
conditions, treatment options and surgical procedures, thus
filling any communication gap between patients and health-
care providers.
BART can be leveraged for named entity recognition
(NER). It can identify and classify specific entities within a
text, such as names of people, medical terms or medications
and aid in organizing and extracting relevant information
from large volumes of text, thus enabling efficient retrieval of
data points. Its ability to comprehend and generate coherent
text aids in performing information extraction tasks. By ana-
lyzing patient data and using the information given, BART
can produce accurate and standardized automated reports
such as radiology, pathology, and laboratory reports. By
leveraging GAI models’ text generation and summarization
capabilities, healthcare organizations can enhance communi-
cation, streamline documentation processes, and improve the
10 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
accessibility and understanding of healthcare information.
However, it is crucial to note that outputs generated by
these models need to be validated by healthcare professionals
to maintain accuracy, reliability, and adherence to ethical
guidelines.
M. VIRTUAL REHABILITATION AND REHABILITATION
ROBOTICS
For a long time, two communities of academics with dif-
fering pensiveness and interests have primarily researched
generative artificial intelligence and virtual environments. In
virtual rehabilitation, where it may help create immersive and
individualized patient experiences, generative AI has much
potential. In order to create interesting and interactive set-
tings for therapeutic reasons, virtual rehabilitation combines
generative models with virtual reality (VR) or augmented
reality (AR) technologies. Because of this pairing, rehabil-
itation patients can have customized immersive experiences.
For example, while movement replication models mimic and
adjust patients’ movements in virtual environments, pro-
viding real-time feedback and encouraging motor control,
gesture recognition models employ generative AI to enable
patients to engage with virtual environments using natural
hand movements. The GAI can also be applied to further
develop and enhance rehabilitation robotics by generating
adaptive robotic movements. The GAI examines patient
features, mobility data, and treatment objectives to create
customized rehabilitation plans. These programs allow robots
to modify their movements, pressures, and levels of support
in response to real-time feedback, maximizing patient par-
ticipation and the efficiency of therapy. Patients can actively
engage in the therapeutic process by using gestures that cause
virtual responses and activities, opening up new paths for
therapeutic interaction. Such interactive components improve
the rehabilitation process and motivate patients, but they also
help patients enhance their motor control abilities and give
real-time feedback on their progression. By using the GAI
in this field, healthcare professionals can facilitate better
recovery outcomes for patients with medical impairments.
N. ADVERSE DRUG REACTION (ADR) PREDICTION
Adverse drug responses (ADRs) are unwanted and adverse
effects of routine drug use. Enhancing drug safety and low-
ering costs can be achieved by anticipating and preventing
ADRs early in drug development. According to a survey,
over 2 million major adverse drug reactions (ADRs) are
estimated to happen among hospitalized patients in the US
each year, resulting in over 100,000 fatalities as a useful tool
for anticipating adverse drug reactions (ADRs) during drug
discovery and development.Generative methods can produce
novel compounds with expected ADR profiles by examining
vast datasets of drug structures, chemical characteristics, and
known ADRs.
In addition to assisting in the early detection and preven-
tion of adverse drug reactions (ADRs), GAI models provide
insightful analyses into the intricate processes that underlie
ADRs, fostering a deeper comprehension of drug safety.
For instance, the BMC Bioinformatics journal published a
neural fingerprint technique in a concurrent deep learning
framework for ADR prediction such that the label infor-
mation (drug-ADR relationship) can be used in the feature
creation stage of the machine learning process. By creat-
ing molecular explanations and emphasizing the pertinent
chemical interactions, generative models can also help us
understand the underlying mechanisms of the ADRs. The
information on pharmaceuticals, ADRs, and target proteins
may be better represented using knowledge graphs and GAI,
which is extremely important for studying ADR prediction.
Using GAI to study ADRs can completely transform drug
discovery and development processes. We can build targeted
therapies and interventions to reduce the impact of ADRs
on patient outcomes as the field develops, and the synergy
of GAI, deep learning, and knowledge graphs improves our
understanding of ADR mechanisms.
O. SYNTHETIC NON-IMAGE DATA
AUGMENTATION/GENERATION
Producing artificial or simulated data that closely mimics ac-
tual data is known as synthetic data production or augmenta-
tion. This method has grown in significance in the healthcare
sector, especially in addressing data privacy issues, extending
datasets, and improving the development and testing of algo-
rithms. The GAI plays a crucial role in developing synthetic
data by utilizing its capability to learn patterns and produce
new data based on existing instances. Generative models
can create artificial data points that capture the statistical
traits and attributes of the original dataset by analyzing vast
amounts of real patient data. This synthetic data can be
utilized alongside current datasets, giving researchers and
developers access to more varied data samples without sacri-
ficing patient privacy. Additionally, controlled and repeatable
datasets can be produced using synthetic data generation for
benchmarking and algorithm evaluation.
The GAI can also generate synthetic data for specific use
cases or scenarios that may be difficult to obtain or rare in
real-world datasets. Techniques like GANs and VAEs have
evolved to generate synthetic tabular data resembling original
datasets, aiding in financial analysis, medical research, and
structured data applications [46].This data can be used in ar-
eas with limited data availability and where generating more
data is difficult or not feasible. Enhancing synthetic data with
the GAI can also reduce the problems associated with imbal-
anced datasets. It is common in the healthcare sector to have
skewed distributions of particular illnesses or demographic
features due to several factors, such as data collection biases
or the rarity of particular diseases. By creating synthetic data
to balance these disparities, generic models can ensure that
algorithms are trained on representative and diverse samples,
leading to more accurate and equitable healthcare solutions.
By effectively utilizing the power of the GAI for synthetic
data augmentation, we can overcome the problem of data
limitation and quicken the creation and implementation of
VOLUME X, 2020 11
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
AI-driven solutions for better patient outcomes and care.
P. BIOMARKER IDENTIFICATION
The process of finding and verifying particular biological
markers that can be utilized as indicators of a particular
disease, physiological condition, or response to treatment is
known as biomarker identification. Biomarkers are quantifi-
able traits that can reveal details about a biological condi-
tion. Examples include chemicals, genes, proteins, or imag-
ing properties [47]. Biomarkers significantly impact disease
diagnosis, prognosis, treatment choice, and monitoring in
medicine and healthcare. Medical experts can learn more
about a disease’s presence, progression, and severity and
assess the efficacy of treatment measures by identifying and
measuring biomarkers [48].
The GAI offers a revolutionary method for identifying
biomarkers in medicine and healthcare. Researchers may
examine enormous and complex biological datasets to find
new biomarkers by utilizing cutting-edge AI models like
Generative Adversarial Networks (GANs) and Variational
Autoencoders (VAEs). The GAI augments conventional sta-
tistical techniques, supplying a more thorough understanding
of biological situations by revealing hidden patterns and link-
ages in genes, proteins, and imaging features. The GAI accel-
erates the detection of biomarkers by creating hypothetical
biomarkers. The biomarker development process can be sped
up, and expenses can be decreased by effectively testing and
validating these AI-generated biomarkers in experimental
settings. Additionally, by producing synthetic data, GAI as-
sists in solving data shortages and privacy issues. In order to
improve the robustness and reliability of biomarker detection,
researchers can expand their datasets by creating synthetic
samples that closely resemble real-world data.
The use of GAI in biomarker identification presents ex-
citing possibilities for improving patient care and medical
research. Artificial intelligence (AI) models speed up the
identification of critical illness indicators and therapeutic
responses by analyzing complex biological data and produc-
ing fictitious biomarkers. This game-changing technology
improves personalized therapy while deepening our under-
standing of various illnesses, opening the door to precision
medicine and better patient outcomes. In essence, while
GAI’s current role in biomarker identification might be in
its infancy, it is untapped potential captivates researchers,
promising groundbreaking strides in understanding biomark-
ers’ role in health and disease. As scientists explore and inno-
vate at the intersection of GAI and biomarker identification,
the prospects for revolutionizing diagnostics, prognostics,
and personalized medicine are incredibly compelling.
Q. DISEASE PROGRESSION MODELING
In disease progression modelling, a disease’s development,
symptoms, and potential effects are examined and predicted
through time. The GAI has the potential to be a powerful
tool in this field and tremendously benefit the healthcare
industry [49]. By analyzing huge databases of patient data,
medical records, and clinical outcomes, generative models
can discover patterns and linkages within the data and imi-
tate disease development scenarios. These models can pro-
duce hypothetical patient trajectories that accurately depict
the development of ailments while considering a variety of
variables, such as genetics, lifestyle, and therapeutic inter-
ventions. This makes it possible for medical practitioners
and researchers to investigate multiple illness trajectories,
assess the efficacy of different interventions, and appreciate
the potential adverse consequences of different treatment
modalities.
Liu et al. [50] introduce a novel end-to-end network
designed to address the complexities of modeling diffuse
gliomas, malignant brain tumors that extensively infiltrate
brain tissue. The intricate interplay between neoplastic
cells, normal tissue, and the changes induced by treatments
presents challenges in accurately modeling glioma tumor
growth. This approach is based on deep-segmentation neural
networks and cutting-edge diffusion probabilistic models to
generate future tumor masks and realistic MRIs depicting the
anticipated tumor appearance at various future time points for
diverse treatment plans. Sequential multi-parametric mag-
netic resonance images (MRI) and treatment information
are used as conditioning inputs to guide the generative dif-
fusion process, enabling tumor growth estimations at any
specific time. By providing treatment-aware generated MRIs,
tumour growth predictions, and uncertainty estimates, the
model offers valuable insights for clinical decision-making,
aiding clinicians in assessing potential outcomes and guiding
treatment strategies.
Also, applying the GAI to progressing illnesses modeling
can enhance clinical judgment, personalized healthcare, and
our comprehension of diseases. In addition to reproduc-
ing disease trajectories, the GAI affects disease progression
modelling. By employing vast amounts of healthcare data
and clinical records, generative models can aid in predict-
ing illness outcomes, evaluating the severity of the disease,
and identifying high-risk patients. To present patients with
customized risk profiles, these models can investigate com-
plex correlations between various data, including the popula-
tion, biomarkers, multiple medical conditions, and lifestyle
choices. This enables medical professionals to make well-
informed decisions about treatment plans and treatments,
leading to more concentrated and effective healthcare prac-
tices. Along with contributing to its predictive abilities, GAI
can support the development of fresh perspectives and dis-
ease mechanism concepts. By examining the created disease
progressions and their related characteristics, researchers can
gain a deeper understanding of the fundamental mechanisms
and processes that underlie the course of disease. This infor-
mation can aid in creating fresh therapeutic targets, make it
simpler to find prospective biomarkers and direct the investi-
gation of novel therapeutic strategies. Moreover, the design
and optimization of clinical trials may be improved using
GAI. Researchers can examine alternative scenarios, evaluate
the effectiveness of various interventions, and calculate the
12 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
potential impact of new therapeutics by modelling illness
development and treatment responses in virtual patient popu-
lations. This can speed up the discovery and approval of new
treatments, lower expenses associated with the clinical trial
process, and streamline the clinical trial process, eventually
benefiting patients by giving them quicker access to cutting-
edge therapies. The modeling of disease progression with
the GAI has enormous potential. Tailored healthcare, better
clinical decision-making, and a greater understanding of dis-
ease mechanisms all benefit from its capacity to analyze big
datasets, simulate disease trajectories, and produce tailored
risk profiles. By utilising the GAI, researchers and healthcare
practitioners can improve treatment plans, understand disease
progress, and boost healthcare delivery.
V. HEALTHCARE-CUSTOMIZED LARGE LANGUAGE
MODELS
Advanced AI systems called large language models (LLMs)
(GAI models) have been trained to understand and produce
language in a way comparable to that of humans. These mod-
els process and analyze text using deep learning techniques,
enabling them to produce consistent and pertinent responses
to the context. large language models offer the potential to
improve human-computer interactions and automate jobs in
numerous sectors where language understanding and gener-
ation are essential. In this section, we describe several GAI
models that are customized for the healthcare domain.
A. MED-PALM
Med-PaLM [51] is a large language model (LLM) created
to offer excellent responses to medical queries. It stands for
Medical Pre-Trained Language Model.It is trained using ex-
tensive medical literature, academic publications, electronic
health records, and other healthcare information. Med-PaLM
has the capacity to comprehend medical jargon, decipher
intricate medical ideas, and produce pertinent comments or
insights [52]. In addition, Med-PaLM [53] produces precise,
beneficial long-form responses to consumer health issues, as
determined by panels of licensed doctors and users. Medical
documentation, electronic health records, medical education
and research, and information retrieval are only some of the
applications that Med-PaLM can be used in the healthcare
industry. Med-PaLM has the potential to improve efficiency,
accuracy, and knowledge availability in numerous areas of
healthcare delivery and research by making use of its ex-
tensive medical knowledge and language-generating capabil-
ities. Recently, Google launched an upgraded model of Med-
PaLM called Med-PaLM 2 [54], which has an 18% leap in
accuracy compared to its predecessor. Med-PaLM 2 achieved
a staggering 86.5% accuracy rate on the United States Med-
ical Licensing Examination (USMLE) questions [52], which
is on par with the "expert" test takers. Figure3 shows the
performance of Med-Palm over other medical models. Med-
Palm 2 could surpass the 60% passing threshold required for
the examination.
FIGURE 3: Performance of Med-PaLM over other healthcare
LLM models
Med-PaLM has several key capabilities that make it valu-
able in the medical sector. For instance, Med-PaLM aids in
the representation of medical knowledge. Anatomy, illnesses,
symptoms, treatments, drugs, and medical procedures are
all included in the Med-PaLM encoding system for medical
knowledge. Because of this expertise, the model can inter-
pret and produce text unique to medical themes. By doing
so, Med-PaLM plays a vital role in representing medical
knowledge. The area of medical documentation is another
one where Med-PaLM is widely employed. Med-PaLM aids
in producing thorough and accurate medical records. It can
construct reports, automatically extract pertinent data from
patient contacts, and help keep standardized terminology
in electronic health records (EHRs). Figure 4 shows the
diagnostic report generated by Med-Palm 2 by analyzing the
image of a chest X-ray [55].
Additionally, Med-PaLM is useful in medical research
and education. Medical students, researchers, and instructors
can use the model’s extensive medical knowledge base and
language-generating skills. To help in studying and com-
prehending difficult medical ideas, Med-PaLM can offer
definitions, justifications, and responses to medical ques-
tions.Condensing research papers, highlighting pertinent in-
formation, and extracting essential conclusions can help with
literature reviews. Med-PaLM may also create hypotheses,
recommend research topics, and promote evidence-based
practice by giving users access to the most recent medical
literature. Information retrieval in the medical industry may
benefit from the use of Med-PaLM. It can efficiently search
for and retrieve pertinent research articles, guidelines, clini-
cal trials, and other sources of medical knowledge because of
its capacity to process and comprehend medical content; This
enables medical practitioners to obtain the most recent, scien-
tifically supported information, fostering informed decision-
making and improving patient care.
Despite the fact that Med-PaLM 2 achieved state-of-the-
VOLUME X, 2020 13
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
art performance on several multiple-choice benchmarks for
medical question answering and that human evaluation shows
answers compare favourably to physician answers across
several clinically important axes, more work needs to be
done to ensure it is used safely and effectively.The ethical
application of this technology will require careful thought, in-
cluding thorough quality assessment when utilized in various
clinical contexts with safeguards to reduce hazards in such
circumstances. For instance, utilizing an LLM to determine a
patient’s diagnosis or course of treatment carries significantly
more risks than using an LLM to learn about a condition or
drug. More research is required to evaluate LLMs used in
healthcare for homogeneity and amplification of biases and
security vulnerabilities inherited from base models. Another
drawback is the potential for Med-PaLM to produce plausible
yet inaccurate or deceptive information. Sometimes language
models can produce responses that appear sensible but lack
medical precision or evidence-based backing. Healthcare
practitioners should use prudence and cross-reference the
data supplied by Med-PaLM with reliable sources and their
own experience. Additionally, Med-PaLM can have trou-
ble handling sensitive patient data and upholding privacy.
When adopting Med-PaLM or any other language model,
appropriate safeguards must be in place to secure sensitive
information because patient privacy and data security are
crucial considerations in the healthcare industry.
B. BIOGPT
BioGPT [56] is a pre-trained transformer language model
that is domain-specifically generative and designed for pro-
ducing and mining biomedical texts. BioGPT is pre-trained
on 15M PubMed abstracts from scratch and adheres to the
transformer language model framework (GPT-2). BioGPT
is capable of carrying out tasks like providing information,
retrieving relevant information, and producing writing perti-
nent to biomedical literature. The goal of BioGPT, which em-
phasizes the biomedical field, is to help researchers, medical
personnel, and scientists with various tasks, such as literature
reviews, drug discovery, protein modelling, and biomedical
data analysis.
In comparison to a single human annotation, BioGPT-
Large scored a record 81% accuracy on PubMedQA. The
accuracy of most other NLP tools, including Google’s BERT
family of language models, has not surpassed that of humans.
BioGPT has several use cases in the field of bio-medicine
and bio-informatics. Personalized medicine, drug discovery,
protein modelling, bio-informatics analysis, literature review,
and educational help in the biomedical and bioinformatics
disciplines are some of the applications of BioGPT. It can
contribute to different facets of scientific research, medical
care, and educational endeavours in these fields because of its
capacity to comprehend and produce human-like language in
the setting of biology. By quickly finding appropriate infor-
mation from voluminous biomedical literature, BioGPT can
support literature reviews and research. Researchers can use
the approach to summarize study articles, extract essential
findings, and detect relationships between various studies;
This helps scholars stay current with the most recent develop-
ments in their fields while saving time. Additionally, BioGPT
can support bioinformatics studies and protein modelling. It
can help with domain identification, protein-protein interac-
tions, and protein structure prediction. BioGPT may provide
researchers with thorough knowledge of genes, pathways,
and biological networks by integrating diverse biological
databases and knowledge sources, which can help with data
interpretation and analysis.
BioGPT presents several of the same challenges as Chat-
GPT despite being trained primarily in biomedical literature.
Inaccurate writing without any references produced by gener-
ative language models is of growing concern because it may
spread false information. Additionally, because BioGPT is
trained using previously published medical studies that may
contain biases, there is a chance that the GAI will reinforce
those prejudices. The model’s ability to produce consistent
and human-like responses raises questions about possible
abuse or the spread of false information. The thorough mon-
itoring, validation, and implementation of suitable controls
to avoid spreading false or biased information are necessary
for the responsible and ethical deployment of BioGPT. While
BioGPT might be useful in decision-making and biological
research, it should not replace the experience and wisdom of
researchers or healthcare professionals. The model’s outputs
should be used to support human judgment rather than as
a replacement for critical thinking, subject-matter expertise,
and expert opinion.
C. IBM WATSON FOR ONCOLOGY
The GAI-powered IBM Watson for Oncology system [57]
was created to aid oncologists in selecting the best course
of treatment for cancer patients. It analyzes a large amount
of medical literature, patient data, and therapy recommenda-
tions using natural language processing, machine learning,
and big data analytics methods [58]. Watson for Oncology
aids oncologists in quickly accessing pertinent medical infor-
mation by processing and comprehending unstructured clin-
ical material. It also offers recommendations for treatments
that are supported by evidence.
Oncologists typically use Watson for Oncology as a deci-
sion support tool in the healthcare sector. It can be connected
with electronic health records (EHRs) and other data to
analyze patient data, including medical history, test results,
pathology reports, and treatment recommendations, sources
[58]. The system evaluates patient information against a
sizable knowledge collection that includes academic papers,
clinical studies, treatment regimens, and medical textbooks.
Then, tailored to each patient’s particular characteristics,
oncologists can receive recommendations and insights on
prospective therapy alternatives. The availability of expert
knowledge is one of IBM Watson for Oncology’s main
advantages [57]. Oncologists can access a lot of clinical
information and medical literature through the system, which
may be difficult to keep up with manually. It provides on-
14 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
cologists with evidence-based treatment options by providing
them with access to the most up-to-date research and clinical
recommendations.
A significant additional benefit is the tailored therapeutic
recommendations offered by Watson for Oncology. The al-
gorithm considers patient-specific characteristics like health
history, genetic data, and treatment response to create per-
sonalized therapeutic options. As a result, oncologists can
tailor treatment plans for specific individuals, considering
their particular needs and circumstances. Additionally, Wat-
son for Oncology has advantages for time management [57].
Oncologists can save a great deal of time by using technology
to quickly assess huge volumes of patient data and medical
literature. Oncologists can review and make wise decisions
more quickly since it concisely delivers synthesized and
pertinent information.
It is important to remember that IBM Watson for Oncology
has some restrictions. The lack of adequate clinical validation
is one drawback. Despite being educated on a vast body of
medical research, Watson for Oncology’s suggestions could
not always coincide with those of particular oncologists or
established institutional policies. The system’s suggestions
need extensive clinical confirmation and be viewed as an
additional tool to aid clinical judgment rather than as a
complete answer [58]. While Watson for Oncology excels
at evaluating vast amounts of data, it might have trouble
deciphering difficult or uncommon instances when there is
a lack of data. It heavily relies on the availability of pertinent
medical literature and clinical evidence to create recommen-
dations, which may be scarcer for particular cancer kinds or
patient populations.
In a broader sense, IBM Watson for Oncology is an
AI-driven system that supports oncologists in treatment
decision-making by offering recommendations supported by
patient data and medical expertise.Personalized treatment
options and access to various information are provided, but
its recommendations should be carefully weighed, and its
incorporation into clinical processes should be rigorously
assessed.
D. HEALTHCARE LANGUAGE MODELS BY NVIDIA
CLARA
The NVIDIA Clara platform is a collection of AI-powered
tools and frameworks created primarily for healthcare ap-
plications [59]. Healthcare language models trained to com-
prehend and interpret content written in medical natural
language are part of the Clara ecosystem [60]. These models
are essential for facilitating efficient clinical natural language
processing (NLP) tasks and assisting healthcare professionals
in making decisions. The substantial medical text data used to
train the NVIDIA Clara healthcare language models includes
electronic health records (EHRs), medical literature, clinical
guidelines, and other medical sources. The models can under-
stand the complex language patterns, medical terminologies,
and context unique to the healthcare sector by utilizing this
enormous corpus of medical material [59]. The deep learning
Can you write a report analyzing this chest X-ray?
Findings
Attached devices: None found
Lungs : No substantial pleural effusion. No
pneumothorax. Lungs appear clear
Cardiomediastinal: Heart size normal. mediastinal
contours in normal limits
Other: No acute skletal abnormality.
Impression:
No active disease seen.. Chest is normal.
FIGURE 4: X-ray report analysis as done by Med-PaLM 2
architectures used to create these language models, including
transformer-based models, have proven incredibly effective
at various tasks involving natural language processing. Clin-
ical entity recognition, medical concept normalization, rela-
tion extraction, and clinical text categorization are just a few
of the NLP tasks the models are taught to handle specifically
for the healthcare industry [60].
Numerous healthcare applications can make use of
NVIDIA Clara’s healthcare language models. They can be
used, for instance, to automatically extract and categorize
clinical entities from unstructured clinical literature, such as
identifying diagnoses, drugs, procedures, and test findings.
This skill is useful for clinical decision assistance, clinical
documentation improvement, and automated medical coding.
The models could potentially improve efforts in tailored
medication. The models can identify significant traits and
trends supporting tailored treatment choices by reading, in-
terpreting, and comprehending clinical narratives distinct to
individual patients; This can help with treatment choice, fore-
casting treatment results, and facilitating precision medicine
techniques. Overall, NVIDIA Clara’s healthcare language
models give the healthcare industry access to the power
of cutting-edge NLP [60]. These models provide insightful
information, boost productivity, and help clinical decision-
making for various healthcare applications by comprehend-
ing and processing medical text data.
VOLUME X, 2020 15
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
E. DEEPHEALTH LLM
MIT and Massachusetts General Hospital researchers cre-
ated the extensive language model known as DeepHealth
[61]. It makes use of machine learning and natural language
processing to meet the specific possibilities and challenges
faced by the healthcare sector. DeepHealth attempts to close
the knowledge gap between unstructured clinical data and
insights that healthcare professionals may use. DeepHealth
is employed in the healthcare sector to improve many facets
of healthcare delivery and decision-making [62]. Answering
clinical questions is a crucial application. In order to de-
liver accurate and contextually appropriate answers to certain
clinical issues posed by healthcare professionals, the model
may assess medical material, such as research articles and
clinical recommendations; This can speed up knowledge
retrieval and aid in the use of evidence when making de-
cisions. Medical image analysis is another field in which
DeepHealth can be applied. The model has been trained to
comprehend radiology reports and medical images, allowing
it to draw out important details and help with complex image
interpretation. It can help with disease diagnosis, abnormality
detection, and insight into treatment planning [63].
By examining patient-specific clinical narratives, Deep-
Health aids in the advancement of personalized medicine
initiatives. The program can process and comprehend each
unique patient’s medical records to extract patterns, risk
factors, and treatment outcomes that help determine the best
course of therapy; This can help with treatment outcome
prediction and precision medicine strategy guidance.
However, DeepHealth has several limitations, much like
any noteworthy language model. The requirement for consid-
erable training data and the potential for biases in the training
data are two major limitations. The calibre and variety of
the data used to train the model significantly impact its
performance. The predictions and suggestions made by the
model could potentially be affected by biases in the data.
Furthermore, it is difficult to interpret the model’s predic-
tions. Even though DeepHealth can offer viewpoints and
suggestions, it can be difficult to comprehend the underlying
logic or describe the decision-making process. Concerns may
be raised by this lack of interpretability in complex medical
situations where explainability is essential.
To summarize, DeepHealth is a large language model
created especially for healthcare applications. In order to
improve clinical question answering, medical picture analy-
sis, and personalized treatment, it uses machine learning and
natural language processing techniques [62]. Benefits include
faster access to medical information, time savings, and en-
couragement of evidence-based decision-making. However,
difficulties with interpretability, domain adaptation, and the
quality of the training data persist, which call for careful
consideration when using the model in various healthcare
contexts.
F. BIOBERT
Developed primarily for biomedical text mining and natural
language processing (NLP) activities in the healthcare sec-
tor, BioBERT (Bidirectional Encoder Representations from
Transformers for Biomedical Text Mining) is a large lan-
guage model. It has been trained on a sizable corpus of
biological literature and is based on the BERT architecture
(Bidirectional Encoder Representations from Transformers)
[64]. BioBERT has shown to be a useful tool for sifting
through biomedical literature and supporting various health-
care applications. BioBERT is frequently used in healthcare
for tasks like text classification, relation extraction, question
answering, and biomedical named entity recognition. It is
extremely good at processing clinical and scientific materials
because it can comprehend the language and vocabulary
unique to the biomedical area. BioBERT can recognize and
extract biomedical items from unstructured text data, includ-
ing diseases, genes, proteins, chemicals, and their interac-
tions [64]. Figure 5 shows how the pre-training and fine-
tuning of large language models like BioBert is done with
the help of the vast medical corpora.
Enhancing the effectiveness and precision of biomedical
information extraction is one of BioBERT’s key advantages.
BioBERT can comprehend the subtleties and intricate link-
ages found in biomedical texts by using the contextual knowl-
edge and semantic representation developed during training;
This helps with activities like automatic annotation, database
curation, and the extraction of insightful information from
research publications [64]. Support for biomedical research
and evidence-based medicine is another benefit of BioBERT.
The model can help with literature reviews, knowledge dis-
covery, and hypothesis development by looking at a lot
of biomedical literature. It gives researchers and healthcare
practitioners access to a vast knowledge base, facilitating
decision-making, study design, and the advancement of sci-
entific knowledge. BioBERT can also fill the divide between
structured and unstructured clinical data. BioBERT enriches
electronic health records (EHRs) and facilitates a more thor-
ough study of patient data by removing pertinent biological
elements and their relationships from clinical narratives [64].
Patient stratification, clinical decision support systems, and
population health analytics can all benefit from this.
However, BioBERT also has certain drawbacks. The avail-
ability and calibre of training data constitute one restriction.
Despite being trained in a vast body of biological literature,
BioBERT may not comprehensively understand some fields
or rare ailments; This may impact its performance on spe-
cialized or understudied issues. Furthermore, biases in the
training data may affect the model’s output and cause unex-
pected biases in applications used afterwards. Additionally,
the computing demands of BioBERT and resource-intensive
nature can make it difficult to implement it on edge devices
or in environments with limited resources. The model’s size
and processing requirements might make it less accessible
and useful in some healthcare applications.
In a nutshell, BioBERT is a potent language model created
16 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Pre-Training Fine-Tuning
Pre-training with bio-
medical corpora
Weight Initializtion
4.5B Words
13.5B Words
Pre-Training Corpora LLM Model Pre-Training Task Specific Datasets LLM Model Fine-Tuning
Question answering
BIOASQ 5a, BIOASQ 5b,....
Named Entity Recognition
BCG2M, NCBI disease,....
Relation Extraction
ChemProt, EU-ADR,....
What is Full form of mTOR
--> Mammalian target of
rapamycin
The adult renal failure cause..
--> O O B I O....
Variants' in @GENES region
contribute to @DISEASES
susceptibility
--> True
FIGURE 5: Pre-Training and Fine-Tuning of Large Language Models
for NLP and biomedical text-mining jobs in the healthcare
sector. Improved information extraction efficiency, support
for research and medicine based on evidence, and enrichment
of clinical data are some of its advantages [64]. However,
while applying BioBERT in healthcare applications, it is
important to take into account its limitations, which include
data accessibility and quality, interpretability, and computing
needs.
G. MED7 LLM
A large language model called Med7 was created especially
for medical NLU (natural language understanding) in the
healthcare sector. Med7 is a clinical text data extraction
tool trained on a wide variety of clinical text data [65].
Unstructured clinical text sources include electronic health
records (EHRs), clinical notes, and medical literature. It
focuses on identifying medical entities and their correspond-
ing qualities to make clinical data analysis more effective
and reliable. Med7 is employed in the healthcare sector
to simplify several processes involving extracting clinical
information. It can automatically identify and extract medical
elements such as diseases, symptoms, medications, therapies,
and test results from unstructured clinical data. Med7 helps
to enhance clinical documentation, improve clinical coding,
and promote clinical decision-making by converting free-text
clinical narratives into structured data [65]. One of Med7’s
key advantages is its better accuracy and efficiency when
processing and understanding clinical language. Applying
the knowledge and context it acquired throughout training
has allowed Med7 to identify medical organizations and
their properties accurately; This improves healthcare infor-
mation’s accuracy and thoroughness while requiring less
manual labour to separate organized data from unorganized
clinical information.
Additionally, Med7 enhances the interoperability and data
integration of healthcare systems. Med7 enables data shar-
ing, communication, and analysis amongst various healthcare
platforms by transforming unstructured clinical language into
structured data [65]; This makes it possible for electronic
health record (EHR) systems, decision support tools, and
other healthcare systems to integrate seamlessly, thereby
increasing patient continuity and data interoperability. The
potential for Med7 to help clinical research and commu-
nity health investigations is an additional advantage. Med7
supports cohort identification, patient stratification, and data
analysis for research by effectively extracting medical enti-
ties and their features from clinical narratives. It makes large-
scale data mining and analytics possible by enhancing both
medical research and evidence-based medicine [65].
However, Med7 also has certain limitations that must be
taken into account. Its reliance on the calibre and variety of
the training data is one of its limitations. The availability and
representativeness of the clinical literature that the model is
trained on can have an impact on how well it performs. The
accuracy and generalizability of the model’s predictions can
be affected by biases or gaps in the training data. Addition-
ally, Med7 acts as a "black box," which can restrict inter-
pretability, like other big language models. Comprehending
the underlying theory or explanation underlying the model’s
predictions can be difficult. In crucial healthcare situations
where explainability is crucial, this lack of interpretability
may cause problems. For proper use of Med7 in healthcare
applications, it is essential to be aware of these constraints.
VI. REAL WORLD PERFORMANCE ASSESSMENT OF
GAI IN HEALTHCARE
Using innovative solutions to overcome traditional methods
is essential in healthcare. The use of GAI has been progres-
sively increasing in various fields. This section will discuss
four real-world use cases where the GAI has played an
essential role in healthcare.
VOLUME X, 2020 17
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
A. VISUAL SNOW SYNDROME
A positive visual disturbance known as visual snow syn-
drome (VSS) [66] has been characterized as the persistent
flickering of countless tiny dots throughout the visual field.
Alternatively, it is similar to viewing the world through the
static noise of an improperly tuned television. The condition
requires innovative approaches for understanding and man-
aging, as its subjective nature and diverse symptomatology
make it challenging to diagnose traditionally. The current
diagnostic criteria for visual snow syndrome include the
presence of minute dots throughout the entire visual field
lasting longer than three months, as well as the presence of at
least two of the visual symptoms listed below: photophobia,
nyctalopia, palinopsia, and entoptic phenomena.
Visual snow is a reasonably uncommon syndrome expe-
rienced worldwide, and as the researchers’ understanding
evolves, exploring new innovative technologies becomes cru-
cial for further understanding and advancement, analysis, and
improving patient care for people suffering from this per-
plexing chronic disease. The diagnosis is primarily based on
patients’ verbal descriptions of their symptoms; this causes
a barrier to understanding the true nature of the disease, as
individual patients may describe their experience differently,
which may be difficult for parents and other healthcare mem-
bers. For better understanding and seeing through the eyes of
the patients, Balas et al. [67] made use of GAI technologies
to generate images of how a patient affected with visual snow
syndrome sees with the help of textual descriptions with the
help of generative artificial intelligence models that could
translate text to the image. Various models were used to get a
clear image, such as DALL·E2 [2], midjourney [3] and stable
diffusion [4]. Figure 6 shows an image generated by Stable
diffusion with the text prompt “seeing a playground through
grainy static interface”. Although the current situation of
generative text-to-image models shows promising results,
more research and training must be done to achieve precise
results that can be directly used in the medical industry.
B. MOLECULAR OPTIMIZATION
Molecular optimization is the process of improving the prop-
erties and characteristics of chemical compounds which can
alter the activity of the target molecule in a desired way. Find-
ing the perfect molecule for a set of specific requirements
organically is very difficult, and designing a molecule from
scratch with all the rightful properties is a challenging and
complex task requiring a lot of time and resources as these
molecular structures have complex properties. Furthermore,
traditional molecule development methods are costly; 2005
pharmaceutical companies spent $2.6 billion to develop new
US Food and Drug Administration-approved drugs [68].
With the growth in computational power and the develop-
ment of new tools, it is essential to utilize advanced AI
tools to optimize drug discovery. Deep generative models are
becoming popular and can automate the generation of new
bioactive and synthesizable molecules.
Molecular optimization can be helped with the help of
FIGURE 6: Generation of an image using stable diffusion
with the text prompt "seeing a playground through grainy
static interface"
generative AI.Molecular Optimization with GAI(MOGA)
evaluates many molecular structures using GAI models. Mol-
CycleGAN [69], a CycleGAN-based model, is a generative
model which improves the compound designing process.
At its core, it has two neural networks: a generator and a
discriminator. The generator network learns to generate real-
istic molecular structures, whereas the discriminator network
tries to distinguish between generated and actual molecules.
Maziarka et al [69]. Discuss how Mol-CycleGAN can gener-
ate a similarly structured molecule with optimized parameter
values. The model was evaluated on optimization objectives
based on structural and physicochemical properties. Octanol-
water partition coefficient (logP) penalized was used to test
molecular optimization. The model was able to increase the
activity of a specific inactive drug.
CADD is Computer Aided Drug Design, which uses in
silico methods to leverage existing chemical knowledge. De
novo design and virtual screening are two main approaches
for drug designing; de novo design uses generative AI models
and has seen rapid progress [70]. Mol-CycleGAN is used to
generate novel molecules which have multiparameter opti-
mization. It leverages image-to-image translation to trans-
form molecular structures between different representations
or chemical spaces, optimizing desired properties. The model
trains from molecules from different chemical spaces named
source space and target space and can map the two spaces
without requiring a direct correspondence between specific
molecules and generate molecules in the target space with
desirable properties. Once the model is trained, it can be used
directly for molecular optimizationby providing the molec-
ular structure from the source space and defining details
from the target space; the model generated has the desired
structure and properties. Figure 7 displays the generation of
the new molecule using GAI models Variational Autoencoder
and MolCycleGAN; the generated molecules have a similar
structure to the original molecule but with optimized penal-
18 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Original
molecule
Reconstructed
molecule
Latent space
Encoder Decoder
MolCycleGAN
Input molecule Generated molecule
Existing property
PlogP: -7.663
Desired property
PlogP: 2.847
FIGURE 7: GAI models for molecular optimization.
ized logP value [69]. As Mol-CycleGAN preserves struc-
tural details, it can maintain cycle consistency; that is, the
generated molecules can be converted back to their original
space while conserving their essential features. It allows the
generated molecules to be refined and improved via iterative
optimization loops.
In their study, Grisoni et al. [71] employed a combina-
tion of generative deep learning models and a microfluidic
platform to achieve a combination of generative AI and on-
chip chemical synthesis and successfully generated liver X
receptor (LXR) agonists. Their research focused on gener-
ating novel molecular compounds using a specifically tuned
computational pipeline that explored the chemical space of
LXR-alpha agonists. The pipeline was limited to products de-
rived from 17 one-step reactions to ensure compatibility with
on-chip synthesis. The GAI model used in the study produced
25 de novo designs, which underwent subsequent retesting,
batch resynthesis and purification. Out of the 14 retested
designs, 12 were confirmed to be potent LXR agonists.
This design-make-test-analyze framework demonstrated its
suitability for drug design purposes. It showcased the po-
tential of combining generative deep-learning models with
microfluidic platforms for efficient and effective chemical
synthesis.
Bagal et al. [72] leverage deep learning techniques, par-
ticularly transformer-based models, for de novo molecule
generation, known as inverse molecular design. This de-
sign utilizes SMILES notation to represent character strings,
enabling natural language processing models for molecular
design. This approach utilizes generative pre-training (GPT)
models to train a transformer-decoder using masked self-
attention for predicting the next token, focusing on generat-
ing druglike molecules. The MolGPT model performs simi-
larly to modern machine learning frameworks in generating
valid, unique, and novel molecules.
GAI models offer their unique ability to generate desired
compounds. This approach allows researchers to navigate
the vast chemical landscape and uncover new compounds
which exhibit many properties, including improved efficacy,
reduced toxicity, enhanced stability, or other desirable char-
acteristics. Utilizing a generative model for molecule synthe-
sis saves time and resources by automating the optimization
process and facilitating the generation of diverse and high-
quality molecular structures.
C. MEDICAL EDUCATION
Generative language models (GLMs) and artificial intelli-
gence (AI) can enhance medical education in various ways,
such as through virtual patients, accurate simulations, cus-
tomized feedback, evaluation techniques, and eliminating
linguistic obstacles. These innovative tools can enhance med-
ical students’ educational outcomes and facilitate immer-
sive learning environments [73]. Interaction between educa-
tors, researchers, and practitioners is essential to developing
healthy practices, regulations, and transparent AI models that
support the moral and responsible use of GLMs and AI in
medical education. Developers can gain greater confidence
and respect from the medical community by being open and
transparent about the data utilized for training, challenges
faced, and evaluation techniques.
Medical education has the potential to be transformed by
contemporary GAI models with enhanced effectiveness, in-
teractivity, and authenticity, such as OpenAI’s ChatGPT and
Google’s BARD [74]. These models provide unprecedented
capabilities, including producing text that sounds like people,
simulating difficult patient scenarios, and delivering cus-
tomized learning experiences [75]; This encourages the cre-
ation of a more interesting and relevant learning environment.
These GAI tools offer a more dynamic and realistic learning
experience than conventional computer-based simulations.
VOLUME X, 2020 19
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
They provide more complex practice settings for medical
students, facilitating clinical judgment and patient care. By
utilizing GLMs’ superior natural language understanding and
producing capabilities, platforms like PerSim [76], a novel
method of delivering medical simulation, provide students
with contextually relevant patient situations that are more
dynamic and adaptive than earlier computer-based models.
During simulation exercises, these AI systems can offer real-
time, personalized feedback based on a learner’s performance
and particular learning needs. Students who receive this eval-
uation can discover their strengths and areas for development.
With the use of GAI in formative and summative assess-
ments, medical education can gain from more individualized,
efficient, and targeted evaluation methods [77]. An example
of the application of GAI in medical education assessments is
the generation of customized quizzes for students. A unique
formative and summative examination can be created for
each student by GAI once their strengths and limitations have
been examined. By analyzing student performance and giv-
ing real-time feedback, these artificial intelligence-driven so-
lutions can assist instructors in creating personalized learning
programs that cater to individual requirements and enhance
overall results. GLMs can be used by a medical educator
to create a variety of simulated patient scenarios. Students
can become familiar with a wide range of medical issues
and patient interactions because of the variety and realism
of these scenarios. For instance, a medical student could
converse with a dummy patient simulating a rare disease, ask
questions, and get answers like actual patients. This can allow
the learner to enhance their clinical reasoning abilities in a
secure setting.
The proficient ability of these models to produce text with
various levels of complexity could improve the availability
of medical information. AI tools can potentially improve the
accessibility and comprehension of health information for a
wide variety of people, from non-specialists to medical pro-
fessionals, by changing the language and terminology used
according to the intended audience. This targeted communi-
cation strategy can increase health literacy and enable people
to make better decisions about their health. By producing
more explanations, examples, and visual aids, language mod-
els like chatGPT can improve medical textbooks. Students’
general comprehension of the subject matter can be improved
by making difficult medical ideas more understandable. Also,
medical students will find it smoother to readily understand a
study’s main conclusions and consequences if language mod-
els like chatGPT are trained to summarize medical research
publications. This can assist students in staying current with
the most recent research in their subject while saving them
time.
While there are numerous current and future potential
benefits of using GAI in the medical education sector, they
are not without a few limitations. Potential problems with
precision, dependability, abuse of AI-generated content, and
worries about academic integrity are real and warrant serious
consideration. The possibility of bias, privacy concerns, and
potential dehumanization in the educational process also
warrant caution. The "digital divide" is a further crucial factor
to consider. Unfair access to AI resources and technology
could worsen existing inequalities in the educational system,
especially in low-resource environments and among disad-
vantaged student groups. In our age, cyberattacks and the
possibility of spreading misinformation are two major wor-
ries that come with AI integration into the medical education
sector.
The medical education area must be particularly watchful
and aggressive in handling these possible issues, given the
high stakes in health care and the potential for harm. So, for
example, the AI-generated material must be of the highest
calibre. It needs to be carefully evaluated to make sure it is
accurate and pertinent. Comprehensive and detailed feedback
loops and appropriate prompting are two strategies that can
help improve the accuracy and dependability of AI-generated
content in medical education. AI systems have been seen to
engage in biased behaviour and further enlarge pre-existing
stereotypes due to their training data and occasionally due
to the skewed dataset. Exercise caution and overcome any
biases when implementing GLMs in medical education.
Numerous earlier examples, such as racial biases in facial
recognition software and Chatbot Tay [78] from Microsoft
tweeting offensive and sexist content, highlight the need for
caution. In addition, ethical and legal concerns are raised by
the use of generative AI in medical education, emphasizing
the necessity for students to get AI ethics training to ensure
the responsible and ethical implementation of these cutting-
edge technologies.
In conclusion, both potential and challenges come with
integrating GLMs and AI into medical education. GAI mod-
els can produce precise, individualized content for pupils,
resulting in more productive learning occasions.It is imper-
ative to properly heed potential biases and ethical problems
to implement these cutting-edge technologies. These mod-
els’ algorithms rely on enormous amounts of data, and if
that data reflects systemic disparities or is biased, it could
perpetuate unequal learning opportunities or strengthen pre-
existing stereotypes. Certain controls and validation proce-
dures must be implemented to ensure justice, equity, and
inclusivity in the instructional content produced by GAI
models. Educators, researchers, and practitioners must work
together to develop standards, regulations, and best prac-
tices that support the moral and efficient integration of GAI
models in medical education. Building trust and credibility
in AI-powered medical education is largely a function of
transparency. Those who developed and implemented these
technologies must openly disclose the underlying algorithms,
data sources, and procedures used in creating educational
content. This openness promotes a better awareness of the
constraints, prejudices, and potential uncertainties related to
AI models, enabling educators and students to assess the
offered instructional content critically.
20 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
D. DENTISTRY
Dental treatment could be enhanced by GAI, which has the
potential to transform the medical field [79]. Dental research
can help to make sure AI is utilized to improve dental
treatment, make it more accessible, and benefit patients, prac-
titioners, and society at large. The introduction of AI in den-
tistry is revolutionizing the industry by enabling higher preci-
sion, fewer mistakes, and reduced human resources needs. By
using contemporary dental technologies like medical robots
and specialized AI models, a new age known as dentistry
might greatly increase dentistry’s reliability, reproducibility,
precision, and efficiency. Additionally, Dentronics might im-
prove risk-assessment techniques, diagnostics, and disease
prediction in addition to improving treatment outcomes by
deepening our understanding of disease pathophysiology.
By examining patient information like CT scans or 3D
models of the mouth and creating the most effective implant
placement plans, GAI models can also help with dental
implant planning.To create individualized treatment plans for
each patient, these models may simulate various scenarios
while considering elements like bone density, nearby teeth,
and functional requirements. Similar to this example, these
GAI models can provide digital photos or 3D models of
potential smile alterations by evaluating facial characteris-
tics, tooth shape, and proportions. Then, dentists can create
customized and realistic smile designs for people seeking
cosmetic dentistry procedures. They can interact with pa-
tients and ensure their needs are properly handled. In order
to identify and categorize different oral disorders, such as
cavities, periodontal diseases, and oral malignancies, GAI
models can be trained to analyze photographs or scans of
the mouth cavity. These models can help dentists make
early diagnoses and detections, improving patient outcomes.
Dental clinical applications and research have a great deal
of potential to be improved by modern language models like
ChatGPT. They can revolutionize dentistry diagnostics and
treatment planning [80].
While GAI models can help in the dental industry, they
should be used to assist dental practitioners rather than taking
the place of practitioners themselves. GAI models should be
viewed as instruments to improve dentists’ talents and the
care they provide for patients because they play a crucial part
in diagnosing and treating oral disorders. Using generative
AI in the dentistry sector comes with limitations as well.
The training data’s precision and variety significantly impact
the correctness and dependability of these models. Therefore,
having incomplete or biased data can result in less-than-
ideal outcomes and could also result in errors in diagno-
sis or treatment planning. The entire clinical context and
patient-specific aspects, such as medical history, lifestyle, or
individual variances, which are critical in dental care, are
not considered by generative AI models. Complex decision-
making in dentistry requires human expertise and judgment,
which AI models cannot fully mimic.
Also, GAI models generally look for patterns and gener-
alizations in training data but could miss important patient-
specific elements. Important factors like medical history,
lifestyle decisions, dental hygiene practices, and individual
differences are not usually properly considered. Because of
this, the produced outputs might not exactly match a patient’s
particular situation, which could result in less-than-ideal
treatment plans or recommendations.Furthermore, GAI mod-
els’ poor interpretability and transparency present a problem.
It becomes more challenging to comprehend the underlying
decision-making process, which limits dentists’ and patients’
capacity to fully trust and comprehend the created outputs.
To make informed decisions and guarantee the safety of their
patients, dental professionals need clear explanations and
clarity about how the AI arrived at its conclusions.
VII. LIMITATIONS OF USING GAI IN HEALTHCARE
The GAI shows great promise in the healthcare industry; it
offers an innovative approach towards traditional methods
and has various advantages. Table 2 shows a list of types of
content forms for which large language models (LLMs) are
available now and possible models which can be available
in the future [18]. However, it is vital to note that GAI has
various limitations, like data bias and ethical considerations.
Understanding these limitations is crucial for advancing this
technology in the healthcare sector. Figure 8 displays the lim-
itations and possible future works that can be implemented
while integrating generative AI in healthcare.
1) Attribution Problem
The difficulty of comprehending and elucidating the moti-
vations behind the judgements or outputs produced by GAI
models in healthcare applications is known as the "attribu-
tion problem" connected with its usage [81]. Deep learning
models, one type of GAI, have demonstrated an extraordinary
ability to produce complicated and realistic data, such as pa-
tient records, medical images, and therapy recommendations.
These models frequently cannot be understood or explained
easily, preventing their responsible and efficient application
in healthcare settings. Because GAI models function as
complicated black boxes, it is challenging to identify and
credit the decision-making process. This poses an attribution
problem. It is more difficult to comprehend how and why
a GAI model arrived at a specific output than conventional
rule-based systems or simpler machine learning models. In
the healthcare industry, this lack of transparency presents
many difficulties. GAI models may unintentionally inherit
biases or unequal representation of specific populations when
trained on huge datasets. Finding and correcting biases in the
generated outputs is difficult without adequate attribution.
This can exacerbate already-existing inequities and result
in differences in healthcare outcomes. Comprehending the
attribution and potential biases in GAI models is crucial to
ensure just and equitable healthcare delivery [81].
The attribution issue also presents ethical and legal issues.
It is essential in the healthcare industry to be able to assign
blame and accountability for choices made. It might be
challenging to pinpoint who or what is at fault if a GAI
VOLUME X, 2020 21
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Type of content Potential applications Availability
Image Analysis Detecting and Analyzing images Yes
Text / Conversations Engaging into human-like conversations
and giving reliable answers to question prompts Yes
Sound Sound based interactions and voice-to-text applications No
Document/PDF Analysis Summarising documents and analyzing research papers No
Video Text to video outputs based on prompts by user Yes
TABLE 2: Types of content forms available that LLMs could analyze now and possible versions in the future
model generates inaccurate or harmful results. This lack of
accountability on a legal and moral level brings concerns
about liability and patient safety. Healthcare organizations
and providers can be reluctant to employ GAI solutions
without clear attribution mechanisms. The attribution issue in
GAI is currently being worked upon. Attention mechanisms,
saliency maps, and post-hoc analysis approaches are just a
few of the methodologies being researched for interpretabil-
ity and comprehensibility. These strategies hope to shed light
on the model’s decision-making process by tying the outputs
to particular features or inputs. Healthcare practitioners can
more accurately evaluate generative AI models’ biases, limi-
tations, and dependability by comprehending the attribution,
ensuring their responsible and efficient usage in healthcare.
2) Contextualization Problem
In the GAI field, contextualization is absorbing and taking
into account pertinent contextual data when producing out-
puts. It entails comprehending the precise context, restric-
tions and needs related to the activity and using that informa-
tion to deliver more precise, pertinent, and significant results.
This includes user preferences, subject-matter expertise,
task-specific requirements, input data, and activity-specific
requirements. Considering these contextual elements, the
model can give results that are more closely aligned with the
specified criteria and appropriate for the specific application
scenario.
Considering the issue’s wider context is another contextu-
alization method. This may consider financial resources, time
restraints, ethical and moral issues, and institutional rules.
GAI models can produce accurate, useful, practical, and
practicable results within the particular application setting by
taking these contextual aspects into account. In health care,
for instance, contextualization would entail adding pertinent
research findings, medical advice, patient-specific data, and
clinical protocols into the generative AI framework. By doing
this, it would be feasible to ensure that the generated outputs,
such as medical diagnoses, treatment recommendations, or
patient records, are accurate, clinically relevant, and in line
with best practices.
The results might not be coherent, relevant, or comply
with particular constraints or rules without sufficient con-
textualization. GAI models are being advanced in terms of
contextualization. Researchers are experimenting with meth-
ods that include utilizing pre-trained models in particular
domains, adding external knowledge bases, focusing on task-
specific data, and creating attention mechanisms that enable
the model to concentrate on pertinent contextual information.
These methods seek to improve the model’s capacity to
produce more accurate, pertinent, and useful outputs across
various applications.
3) Data quality and bias
One of the significant limitations of GAI is the data quality
and bias. The framework of GAI produces output based on
the training data provided by health care records and medical
writings; if the data supplied is biased or of poor quality, it
will lead to inaccurate results. These data biases can creep
into any development stage and come from several factors,
including demographic disparities, variations in healthcare
practices, and under-representation of specific populations.
The GAI model can amplify these biases and perpetuate
healthcare disparities and unequal treatment outcomes. Many
generative AI applications in public deployment have al-
ready seen neglect in addressing bias. The data is unseeingly
scraped from the internet without giving much attention
to potential sources of misinformation or bias [82] or the
training data is generated through unreliable sources [83].
Although responses from chatbots powered by LLM may
seem creative, they simply reflect the model’s extensive ana-
lytical understanding of which words have been used before
others in the text that it has already viewed. They cannot
understand any language they use, including their responses
and the prompts they are given. Models trained on a large
body of internet data with little filtering (such as ChatGPT or
steady diffusion, for example) have absorbed facts and false
information, biased and fair stuff, and hazardous and innocu-
ous things. LLMs run the risk of duplicating, amplifying, and
spreading bad content and false information without a way to
evaluate any of these characteristics prior to responding to a
prompt, as shown by many examples.
4) Generalization of unseen data
The generalization of unseen data is essential for GAI models
to be reliable in real-world healthcare systems. However, this
generalization can be affected due to various factors. If the
generative AI models fail to update data over time or adapt
to new medical trends, they will struggle to generalize to the
current data distribution. Furthermore, the generative models
are sensitive to variations and noise in the input data. This
22 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
noise and variations can occur due to measurement errors
or differences in imaging techniques. If the model cannot
handle these issues, it will face performance issues when
dealing with new data, which is different from training data.
Furthermore, in healthcare, choosing the correct training data
without infringement of copyrights or other ethical consid-
erations while opting for optimum model performance is
challenging.
5) Patient data privacy
Patient data privacy is one of the major concerns in health-
care. In healthcare treatment, ambient sensors will collect
patient data, including name, age, area of residence, medical
history and other relevant information. The sensors may also
capture information like the patient’s voice, face, or heart
rate, depending on the hardware. Information like this, if
leaked, could lead to the exposure of patients’ health status
and their private information. [84]
Furthermore, many patients receive in-house treatment or
healthcare facilities considered free from sensors.For exam-
ple, a patient might want to restrict monitoring a particular
body part during a particular duration while using the bath-
room. GAI models must adhere to this decisional privacy [85]
and give the right to decide on their privacy and the amount
of information they want to share.
The United States has data protection regulation- HIPAA
to safeguard patient privacy. Generative AI models can be
vulnerable to adversarial attacks. Malicious hackers can try
to manipulate the models’ outputs or get unauthorized access
to sensitive information [86]. Countersecurity measures such
as intrusion detection systems, encryption, and authentication
must be implemented to avoid such attacks.
6) False information generation
An LLM might occasionally reveal the truth or create infor-
mation that is pertinent, acceptable, occasionally surprising,
innovative, and appealing. Other times, it might generate or
support the most flagrant and harmful falsehood. Second,
the model cannot determine which one it is at any given
moment, let alone alert the user. It is unsure whether the
stuff it creates tells the truth or contains fabrications, misrep-
resentations, or objectionable material. Moreover, because
LLMs are probabilistic algorithms, they may return different
responses when given the same task or question more than
once. These responses may be updated versions of previously
incorrect or complicated answers, updated versions of incor-
rect answers that were previously correct, or combinations
of these. This behaviour creates an issue with repeatability
and reliability that necessitates ongoing human supervision
of model operation.
7) Integration with current healthcare technologies
Implementing generative AI models into the current health-
care systems can be challenging. The generative models must
be compatible with electronic health record (EHR) systems,
clinical decision support tools, and other healthcare IT infras-
tructure. This integration can be complex and may require
further study.To access real-time scenarios, generative AI
models should be capable of processing data and generating
outputs in real-time to ensure that healthcare professionals
can make timely decisions. This real-time processing re-
quires efficient computational infrastructure and optimized
algorithms, which must be programmed appropriately before
implementation.
8) Computational Cost
One significant factor to consider while integrating GAI into
healthcare is cost. While implementing GAI is revolutionary,
it may take a toll on the pocket. Training and building gener-
ative AI models require significant resources, expertise, and
computational infrastructure investment. It requires hiring
AI technicians, investing in powerful hardware and software
and acquiring large datasets for models to train on. While
some resources require only a bigger initial payment, others
supply recurring bills. These include upgrading the models,
maintaining the hardware and addressing potential security
vulnerabilities.
Furthermore, Healthcare professionals must be trained to
integrate this technology into their daily dealings. Providing
comprehensive training programs will incur costs associated
with building a new curriculum, training sessions, and con-
tinuous learning initiatives.
9) Lack of professional expertise
It is important to note that generative AI models cannot
replace professional human involvement. While the models
can assist in decision-making, the responsibility of patient
care must be taken by healthcare professionals, and recom-
mendations by these models must be considered after proper
validation by human clinicians. Guidelines and regulatory
frameworks to address these ethical considerations must be
established to govern the deployment and use of GAI models
in healthcare.
10) Ethics
Ethics plays a pivotal role in ensuring the responsibility
and deployment of generative AI in healthcare, and the
integration of AI into healthcare has introduced complex
ethical considerations which must be taken into care to
uphold patient privacy, fairness, accountability and overall
well-being of the individual. Legal and ethical frameworks
are mandatory to establish to prevent privacy concerns. Strict
laws and governing bodies may protect a patient’s overall
treatment process. A recent high-profile lawsuit on Stable
diffusion reveals the use of "derivative works" by OpenAI’s
models [87]. It suggests that LLM models use their work as
a reference to generate images without the artist’s consent,
thus invoking copyright issues. These issues, if not dealt with,
can affect patient treatment and harm the patient in the long
run as these practices will not be limited to copyrights but
may also extend to the patient’s consent. One major factor in
VOLUME X, 2020 23
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Regulatory challenges Description
Intellectual Property Intellectual Property issues can be generated if GAI models create proprietary
medical literature or research.
Medical Malpractice Liability
If recommendations given by GAI models bring harm to the patient when
implemented, who should be held accountable? Healthcare professionals who
used it, model engineers, or the institution that granted permission for usage.
Quality Control & Standardization Consistency and reliability of recommendations made by GAI models
need to be regulated as well as the data which is used to train the model.
Data Ownership It is hard to justify who owns the data from which the LLMs learn.
Concerns are raised when it comes to patients’ data.
Continuous Monitoring & Validation Ensuring continuous accuracy, performance and validity of GAI models
over different categories at all times is a challenging task.
Informed Consent
It can be difficult for patients to understand the implications of GAI usage
in their treatment; in every case, GAI is used, the patient must be informed
of its pros and cons.
Interpretability & Transparency
Transparency must be maintained as to how and why a GAI model suggested
a particular treatment. It can be difficult to explain every step taken by GAI
model to patients.
Over-reliance on GAI Models Over-reliance on GAI models can limit the need to consult professionals every
now and then. It can lead to serious implications if the model malfunctions.
TABLE 3: Regulatory Challenges in Implementation of Generative AI in Healthcare
ethical considerations is transparency. In the current scenario,
a user cannot tell if the output provided by GAI models on
user prompts is true without any external assistance. It has
been noted that some GAI models, such as Galactica AI,
displayed made-up citations and papers [88]. If such models
are used without proper checking, it will lead to misdiagnosis
and fatal treatment. In addition to these, Table 3 [18] lists
various regulatory challenges that must be considered when
implementing generative AI in healthcare [82].
Apart from the above challenges, there is also a chance
that excessive dependence on LLMs will result in the un-
dervaluation of clinical judgment and human competence.
While LLMs can support decision-making, they should not
replace a healthcare professional’s knowledge, experience,
or capacity for critical thought. It is crucial to balance
the capabilities of LLMs and the engagement of human
professionals to maintain patient safety and the best possi-
ble healthcare outcomes. GAI researchers, medical experts,
ethicists, and regulatory agencies must continue to work
together on research, development, and collaborative projects
to address these limitations. To maximize the advantages
of LLMs while minimizing their drawbacks and potential
risks in future healthcare applications, transparent standards,
strong validation processes, and strict ethical frameworks
are essential. These should be seen as inadequate machines
that have the potential to greatly increase process efficiency
but necessitate tight human oversight and intervention at all
operational interfaces, including input and output.
VIII. FUTURE RESEARCH DIRECTIONS
GAI applications that create new content in response to
textual instructions, including text, images, audio, code, and
videos, heavily rely on large language models. These GAI
applications can become techniques with significant potential
for spreading false information or damaging and erroneous
content at an unprecedented scale without human oversight,
guidance, and responsible design and operation. They could,
however, develop into extremely effective, reliable aids for
information management provided they are positioned and
developed responsibly as companions in offering support to
people, enhancing but not replacing their role in decision-
making, knowledge retrieval, and other cognitive processes.
Enhancing these models’ clinical decision-support skills is
becoming increasingly a priority. In this section, we present
future directions for research in GAI for healthcare.
A. CUSTOMIZED/PERSONALIZED SUGGESTIONS AND
A PLATFORM FOR INFORMATION EXCHANGE
LLMs can be progressively improved to provide more pre-
cise and individualized suggestions for diagnosis, therapy
planning, and patient outcome monitoring. These models
can deliver timely and context-aware insights to healthcare
practitioners, improving clinical decision-making. They in-
corporate real-time patient data, such as electronic health
records and information from wearable devices. LLMs may
also help professionals in the healthcare industry collaborate
across disciplines and share knowledge. These models can
work as a shared platform for information exchange, enabling
practitioners from different disciplines to interact and gain
from each other’s experience. They are accessible to many
practitioners, including doctors, nurses, and allied healthcare
professionals. This partnership may result in more thorough
and all-encompassing patient care and a culture of ongoing
learning and development among healthcare professionals.
24 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
Limitations and Future Works for
Using GAI in Healthcare
Data Quality and
Bias
Patient Data
Privacy
Integration with
Health System Ethics
Personalized
Platform
Better Patient
Worker Interaction
Streamlining
Operations
Bridging
Knowledge Gap
Attribution
Problem
Contextualization
Problem
Generalization of
unseen data
Lack of
professional
expertise
FIGURE 8: Limitations and Future Work
B. ENHANCED PATIENT AND WORKER INTERACTIONS
Enhancing LLMs’ natural language comprehension abilities
to understand better medical jargon, contextual nuances, and
patient-specific information is another topic of future efforts.
This would allow patients and virtual assistants powered by
these models to interact more successfully. Patients could
provide precise information about their medical conditions,
receive explanations and recommendations specific to their
needs, and even have dialogues that resemble human inter-
actions. This could enhance patient education, involvement,
and informational access to healthcare. LLMs could also
help with cross-language communication and multilingual
healthcare encounters. These models’ ability to translate
between languages can assist healthcare workers and pa-
tients who speak different languages to communicate more
effectively. Ensuring accurate and effective communication
between patients and healthcare practitioners might consid-
erably increase access to healthcare services, especially in
multicultural and diverse areas.
C. STREAMLINING ADMINISTRATIVE OPERATIONS
LLMs can also help the healthcare sector’s administrative
operations run more smoothly. They can streamline and
automate tasks like invoicing, coding, and medical docu-
mentation, giving healthcare workers more time to provide
direct patient care. These models can help create precise
and uniform clinical notes, summarise patient encounters,
and extract pertinent data from medical records, improving
efficiency and easing administrative stress.LLMs may one
day be essential to advancing medical science. They can help
with information synthesis, literature reviews, and finding
patterns or relationships in the massive body of biological
literature. These models could help with drug discovery
and repurposing initiatives, advancing personalized medical
techniques, and accelerating scientific advancements.
D. ENHANCING DECISION MAKING AND BRIDGING
THE KNOWLEDGE GAP
LLMs can also close the time gap between the limited time
available to healthcare professionals and the continually de-
veloping body of medical knowledge. Keeping up with the
most recent research can be difficult for busy practitioners
due to the constant influx of new research papers, clinical
trials, and treatment guidelines. On the other hand, LLMs
may continuously study and analyze the most recent data,
guaranteeing that healthcare professionals have access to the
most up-to-date and pertinent insights. This modern infor-
mation integration raises the standard of care by enabling
practitioners to make educated judgments based on the most
recent data.
IX. CONCLUSION
GAI models, recently catching attention, have promising
potential in healthcare. This paper has discussed various
applications in which different GAI models enhanced health-
care operations. The diverse applications, including medi-
cal imaging, drug discovery, personalized patient treatment,
medical simulation and training, clinical trial optimization,
mental health support, healthcare operations and resource
management, chatbots, human movement simulation and
analysis and text generation and summarization indicate how
flexible and reliable this technology is and how it can be
VOLUME X, 2020 25
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
implemented in healthcare. The different use cases discussed
show how GAI has been used to aid patients suffering from
visual snow syndrome and enhance the molecular optimisa-
tion process. Further, how GAI is used in Medical Education
and Dentistry has also been showcased. However, despite
the gains, there continue to be challenges associated with
applying GAI in healthcare. Ethics, including patient privacy
and data security, must be prioritized. Stringent laws and
safety measures must be in place to ensure the appropriate
and secure use of patient information. More study and anal-
ysis are required to utilise GAI’s power in regular health-
care practices properly. Collaboration between AI scientists,
healthcare practitioners, and legislators is essential to solve
technological constraints, validate models and smoothly in-
tegrate GAI into current healthcare systems.
REFERENCES
[1] L. Floridi and M. Chiriatti, “Gpt-3: Its nature, scope, limits, and conse-
quences,” Minds and Machines, vol. 30, pp. 681–694, 2020.
[2] G. Marcus, E. Davis, and S. Aaronson, “A very preliminary analysis of
dall-e 2,” arXiv preprint arXiv:2204.13807, 2022.
[3] “Midjourney.” [Online]. Available: https://www.midjourney.com/home/
?callbackUrl=%2Fapp%2F
[4] “Stable diffusion.” [Online]. Available: https://stablediffusionweb.com/
[5] J. Zhang, Y. Peng, and M. Yuan, “Unsupervised generative adversarial
cross-modal hashing,” in Proceedings of the AAAI conference on artificial
intelligence, vol. 32, no. 1, 2018.
[6] D. Baidoo-Anu and L. O. Ansah, “Education in the era of generative
artificial intelligence (ai): Understanding the potential benefits of chatgpt
in promoting teaching and learning,” Journal of AI, vol. 7, no. 1, pp. 52–62,
2023.
[7] O. Temsah, S. A. Khan, Y. Chaiah, A. Senjab, K. Alhasan, A. Jamal,
F. Aljamaan, K. H. Malki, R. Halwani, J. A. Al-Tawfiq et al., “Overview
of early chatgpt’s presence in medical literature: insights from a hybrid
literature review by chatgpt and human experts,” Cureus, vol. 15, no. 4,
2023.
[8] “Diagnamed launches a generative artificial intelligence medical chatbot,
dr. genai,” https://aithority.com/machine-learning/diagnamed-launches-
a-generative-artificial- intelligence-medical- chatbot-dr-genai/, (Accessed
on 06/19/2023).
[9] M. Sabry Abdel-Messih and M. N. Kamel Boulos, “Chatgpt in clinical
toxicology,” JMIR Medical Education, vol. 9, p. e46876, 2023.
[10] M. Sallam, “Chatgpt utility in healthcare education, research, and practice:
systematic review on the promising perspectives and valid concerns,” in
Healthcare, vol. 11, no. 6. MDPI, 2023, p. 887.
[11] T. N. Y. Times, “Generative a.i. start-up cohere valued at about $2 billion
in funding round,” 2023. [Online]. Available: https://www.nytimes.com/
2023/05/02/technology/generative-ai-start- up-coherefunding.html
[12] R. Mao, G. Chen, X. Zhang, F. Guerin, and E. Cambria, “Gpteval: A survey
on assessments of chatgpt and gpt-4,” arXiv preprint arXiv:2308.12488,
2023.
[13] T. H. Kung, M. Cheatham, A. Medenilla, C. Sillos, L. De Leon,
C. Elepaño, M. Madriaga, R. Aggabao, G. Diaz-Candido, J. Maningo
et al., “Performance of chatgpt on usmle: Potential for ai-assisted medical
education using large language models,” PLoS digital health, vol. 2, no. 2,
p. e0000198, 2023.
[14] O. AI, “Gpt-4 is openai’s most advanced system, producing safer and more
useful responses,” 2023, https://openai.com/gpt-4.
[15] S. Teebagy, L. Colwell, E. Wood, A. Yaghy, and M. Faustina, “Improved
performance of chatgpt-4 on the okap exam: A comparative study with
chatgpt-3.5,” medRxiv, pp. 2023–04, 2023.
[16] G. A. Johnson, J. N. Bloom, L. Szczotka-Flynn, D. Zauner, and R. L.
Tomsak, “A comparative study of resident performance on standardized
training examinations and the american board of ophthalmology written
examination,” Ophthalmology, vol. 117, no. 12, pp. 2435–2439, 2010.
[17] R. Bhayana, S. Krishna, and R. R. Bleakney, “Performance of chatgpt on
a radiology board-style examination: Insights into current strengths and
limitations,” Radiology, p. 230582, 2023.
[18] B. Meskó and E. J. Topol, “The imperative for regulatory oversight of large
language models (or generative ai) in healthcare,” npj Digital Medicine,
vol. 6, no. 1, p. 120, 2023.
[19] R. Rothbaum and J. S. McGee, “Aquagenic urticaria: diagnostic and
management challenges,” Journal of asthma and allergy, pp. 209–213,
2016.
[20] J. T. Ludlow, R. G. Wilkerson, and T. M. Nappe, “Methemoglobinemia,”
2019.
[21] Y. Yang, H. Fu, A. I. Aviles-Rivero, C.-B. Schönlieb, and L. Zhu, “Diffmic:
Dual-guidance diffusion network for medical image classification,” in
International Conference on Medical Image Computing and Computer-
Assisted Intervention. Springer, 2023, pp. 95–105.
[22] X. Bing, W. Zhang, L. Zheng, and Y. Zhang, “Medical image super
resolution using improved generative adversarial networks,” IEEE Access,
vol. 7, pp. 145 030–145 038, 2019.
[23] J. S. Yoon, C. Zhang, H.-I. Suk, J. Guo, and X. Li, “Sadm: Sequence-
aware diffusion model for longitudinal medical image generation,” in
International Conference on Information Processing in Medical Imaging.
Springer, 2023, pp. 388–400.
[24] T. Susnjak, “Beyond predictive learning analytics modelling and onto
explainable artificial intelligence with prescriptive analytics and chatgpt,”
International Journal of Artificial Intelligence in Education, pp. 1–31,
2023.
[25] C. Molnar, G. Casalicchio, and B. Bischl, “Interpretable machine learning–
a brief history, state-of-the-art and challenges,” in Joint European con-
ference on machine learning and knowledge discovery in databases.
Springer, 2020, pp. 417–431.
[26] M. T. Ribeiro, S. Singh, and C. Guestrin, “Anchors: High-precision model-
agnostic explanations,” in Proceedings of the AAAI conference on artifi-
cial intelligence, vol. 32, no. 1, 2018.
[27] Y. Li, J. Pei, and L. Lai, “Structure-based de novo drug design using 3d
deep generative models,” Chemical science, vol. 12, no. 41, pp. 13 664–
13 675, 2021.
[28] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and
X. Chen, “Improved techniques for training gans,” Advances in neural
information processing systems, vol. 29, 2016.
[29] H. Abdine, M. Chatzianastasis, C. Bouyioukos, and M. Vazirgiannis,
“Prot2text: Multimodal protein’s function generation with gnns and trans-
formers,” arXiv preprint arXiv:2307.14367, 2023.
[30] C. Buche, F. Lasson, and S. Kerdelo, “Conditional autoencoder pre-
training and optimization algorithms for personalized care of hemophiliac
patients,” Frontiers in Artificial Intelligence, vol. 6, 2023.
[31] H. Luo, K. Nagano, H.-W.Kung, Q. Xu, Z. Wang, L. Wei, L. Hu, and H. Li,
“Normalized avatar synthesis using stylegan and perceptual refinement,” in
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition, 2021, pp. 11 662–11672.
[32] J. Mueller, R. B. Parikh, and A. Noble, “Evaluating clinical trial inclu-
sion/exclusion criteria from claims using generative artificial intelligence.”
2023.
[33] L. Gootjes-Dreesbach, M. Sood, A. Sahay, M. Hofmann-Apitius, and
H. Fröhlich, “Variational autoencoder modular bayesian networks for
simulation of heterogeneous clinical study data,” Frontiers in big Data,
vol. 3, p. 16, 2020.
[34] L. Yang, “Multi-modal depression detection and estimation,” in 2019 8th
International Conference on Affective Computing and Intelligent Interac-
tion Workshops and Demos (ACIIW). IEEE, 2019, pp. 26–30.
[35] “Revolutionizing mental health: Generative ai in therapy,”
https://www.productiveedge.com/blog/revolutionizing-mental-health-
generative-ai-and- therapy, (Accessed on 12/11/2023).
[36] “Generative ai for mental health is upping the ante by going
multi-modal, embracing e-wearables, and a whole lot more,”
https://www.forbes.com/sites/lanceeliot/2023/11/02/generative-ai- for-
mental-health- is-upping-the-ante-by- going-multi-modal-embracing-e-
wearables-and- a-whole-lot-more/, (Accessed on 12/11/2023).
[37] M. Barat, P. Soyer, and A. Dohan, “Appropriateness of recommendations
provided by chatgpt to interventional radiologists,” Canadian Association
of Radiologists Journal, p. 08465371231170133, 2023.
[38] T. Dave, S. A. Athaluri, and S. Singh, “Chatgpt in medicine: an overview
of its applications, advantages, limitations, future prospects, and ethical
considerations,” Frontiers in Artificial Intelligence, vol. 6, 2023.
[39] H. Lee, “The rise of chatgpt: Exploring its potential in medical education,”
Anatomical Sciences Education, 2023.
26 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
[40] “What is instructgpt and how to access instruct gpt?” https://www.
theinsaneapp.com/2023/05/everything-about-instructgpt.html, (Accessed
on 06/29/2023).
[41] “Ai ushers in next-gen prior authorization in healthcare | mckinsey |
mckinsey,” https://www.mckinsey.com/industries/healthcare/our-insights/
ai-ushers- in-next-gen- prior-authorization-in-healthcare, (Accessed on
07/17/2023).
[42] “Triage - an overview | sciencedirect topics,” https://www.sciencedirect.
com/topics/medicine-and- dentistry/triage, (Accessed on 07/17/2023).
[43] J. Huang, L. Neill, M. Wittbrodt, D. Melnick, M. Klug, M. Thompson,
J. Bailitz, T. Loftus, S. Malik, A. Phull et al., “Generative artificial intel-
ligence for chest radiograph interpretation in the emergency department,”
JAMA network open, vol. 6, no. 10, pp. e2336 100–e2 336100, 2023.
[44] “Safely harness the power of generative ai in healthcare with
clinically validated virtual triage,” https://www.prnewswire.com/news-
releases/safely-harness- the-power-of- generative-ai-in-healthcare-
with-clinically- validated-virtual-triage- 301811636.html, (Accessed on
07/17/2023).
[45] D. M. Levine, R. Tuwani, B. Kompa, A. Varma, S. G. Finlayson,
A. Mehrotra, and A. Beam, “The diagnostic and triage accuracy of the
gpt-3 artificial intelligence model,” medRxiv, pp. 2023–01, 2023.
[46] A. Jadon and S. Kumar, “Leveraging generative ai models for synthetic
data generation in healthcare: Balancing research and privacy,” arXiv
preprint arXiv:2305.05247, 2023.
[47] H. Mischak, G. Allmaier, R. Apweiler, T. Attwood, M. Baumann, A. Be-
nigni, S. E. Bennett, R. Bischoff, E. Bongcam-Rudloff, G. Capasso et al.,
“Recommendations for biomarker identification and qualification in clini-
cal proteomics,” Science translational medicine, vol. 2, no. 46, pp. 46ps42–
46ps42, 2010.
[48] R. Nair, D. D. Mohan, S. Frank, S. Setlur, V. Govindaraju, and
M. Ramanathan, “Generative adversarial networks for modelling clinical
biomarker profiles with race/ethnicity,” British Journal of Clinical Phar-
macology, vol. 89, no. 5, pp. 1588–1600, 2023.
[49] O. Zaballa, A. Pérez, E. G. Inhiesto, T. A. Ayesta, and J. A. Lozano,
“Learning the progression patterns of treatments using a probabilistic
generative model,” Journal of Biomedical Informatics, vol. 137, p. 104271,
2023.
[50] Q. Liu, E. Fuster-Garcia, I. T. Hovden, D. Sederevicius, K. Skogen,
B. J. MacIntosh, E. Grødem, T. Schellhorn, P. Brandal, A. Bjørnerud
et al., “Treatment-aware diffusion probabilistic model for longitudinal
mri generation and diffuse glioma growth prediction,” arXiv preprint
arXiv:2309.05406, 2023.
[51] K. Singhal, S. Azizi, T. Tu, S. S. Mahdavi, J. Wei, H. W. Chung, N. Scales,
A. Tanwani, H. Cole-Lewis, S. Pfohl et al., “Large language models
encode clinical knowledge,” arXiv preprint arXiv:2212.13138, 2022.
[52] K. Singhal, T. Tu, J. Gottweis, R. Sayres, E. Wulczyn, L. Hou, K. Clark,
S. Pfohl, H. Cole-Lewis, D. Neal et al., “Towards expert-level med-
ical question answering with large language models,” arXiv preprint
arXiv:2305.09617, 2023.
[53] H. Bair and J. Norden, “Large language models and their implications on
medical education,” Academic Medicine, pp. 10–1097, 2023.
[54] “Health ai research llm updates,” https://blog.google/technology/health/ai-
llm-medpalm- research-thecheckup/, (Accessed on 06/28/2023).
[55] “Med-palm,” https://sites.research.google/med-palm/, (Accessed on
07/14/2023).
[56] R. Luo, L. Sun, Y. Xia, T. Qin, S. Zhang, H. Poon, and T.-Y. Liu,
“Biogpt: generative pre-trained transformer for biomedical text generation
and mining,” Briefings in Bioinformatics, vol. 23, no. 6, 2022.
[57] Z. Jie, Z. Zhiying, and L. Li, “A meta-analysis of watson for oncology in
clinical application,” Scientific reports, vol. 11, no. 1, p. 5792, 2021.
[58] Z. Dlamini, F. Z. Francies, R. Hull, and R. Marima, “Artificial intelligence
(ai) and big data in cancer and precision oncology,” Computational and
structural biotechnology journal, vol. 18, pp. 2300–2311, 2020.
[59] “Intelligent computing for healthcare | nvidia clara,” https://www.nvidia.
com/en-in/clara/, (Accessed on 07/17/2023).
[60] “Healthcare developer resources | nvidia developer,” https://developer.
nvidia.com/industries/healthcare, (Accessed on 07/17/2023).
[61] “Deep health | radiology ai machine learning solutions,” https://deephealth.
com/, (Accessed on 07/17/2023).
[62] “Deephealth: A deep learning solution for the covid-19 crisis | by pablo
castañeda | saturdays.ai | medium,” https://medium.com/saturdays-
ai/deephealth-a- deep-learning-solution-for-the-covid-19-crisis-
785238119c1a, (Accessed on 07/17/2023).
[63] A. S. Panayides, A. Amini, N. D. Filipovic, A. Sharma, S. A. Tsaftaris,
A. Young, D. Foran, N. Do, S. Golemati, T. Kurc et al., “Ai in medical
imaging informatics: current challenges and future directions,” IEEE jour-
nal of biomedical and health informatics, vol. 24, no. 7, pp. 1837–1857,
2020.
[64] J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C. H. So, and J. Kang, “Biobert:
a pre-trained biomedical language representation model for biomedical
text mining,” Bioinformatics, vol. 36, no. 4, pp. 1234–1240, 2020.
[65] A. Kormilitzin, N. Vaci, Q. Liu, and A. Nevado-Holgado, “Med7: A
transferable clinical natural language processing model for electronic
health records,” Artificial Intelligence in Medicine, vol. 118, p. 102086,
2021.
[66] F. Puledda, C. Schankin, K. Digre, and P. J. Goadsby, “Visual snow
syndrome: what we know so far,” Current opinion in neurology, vol. 31,
no. 1, pp. 52–58, 2018.
[67] M. Balas and J. A. Micieli, “Visual snow syndrome: Use of text-to-image
artificial intelligence models to improve the patient perspective,” Canadian
Journal of Neurological Sciences, pp. 1–2, 2022.
[68] J. Avorn, “The $2.6 billion pill—methodologic and policy considerations,”
New England Journal of Medicine, vol. 372, no. 20, pp. 1877–1879, 2015.
[69] Ł. Maziarka, A. Pocha, J. Kaczmarczyk, K. Rataj, T. Danel, and M. War-
choł, “Mol-cyclegan: a generative model for molecular optimization,”
Journal of Cheminformatics, vol. 12, no. 1, pp. 1–18, 2020.
[70] M. H. Segler, T. Kogej, C. Tyrchan, and M. P. Waller, “Generating focused
molecule libraries for drug discovery with recurrent neural networks,”
ACS central science, vol. 4, no. 1, pp. 120–131, 2018.
[71] F. Grisoni, B. J. Huisman, A. L. Button, M. Moret, K. Atz, D. Merk, and
G. Schneider, “Combining generative artificial intelligence and on-chip
synthesis for de novo drug design,” Science Advances, vol. 7, no. 24, p.
eabg3338, 2021.
[72] V. Bagal, R. Aggarwal, P. Vinod, and U. D. Priyakumar, “Molgpt: molec-
ular generation using a transformer-decoder model,” Journal of Chemical
Information and Modeling, vol. 62, no. 9, pp. 2064–2076, 2021.
[73] J. G. Ruiz, M. J. Mintzer, and R. M. Leipzig, “The impact of e-learning
in medical education,” Academic medicine, vol. 81, no. 3, pp. 207–212,
2006.
[74] M. Karabacak, B. B. Ozkara, K. Margetis, M. Wintermark, and S. Bisdas,
“The advent of generative language models in medical education,” JMIR
Medical Education, vol. 9, p. e48163, 2023.
[75] Y. Okuda, E. O. Bryson, S. DeMaria Jr, L. Jacobson, J. Quinones,
B. Shen, and A. I. Levine, “The utility of simulation in medical education:
what is the evidence?” Mount Sinai Journal of Medicine: A Journal of
Translational and Personalized Medicine: A Journal of Translational and
Personalized Medicine, vol. 76, no. 4, pp. 330–343, 2009.
[76] “What is persim®? - medcognition,” https://medcognition.com/what-is-
persim/, (Accessed on 06/29/2023).
[77] “20230624-22994-cporbe.pdf,” https://assets.cureus.com/uploads/
review_article/pdf/158756/20230624-22994-cporbe.pdf, (Accessed
on 07/14/2023).
[78] “In 2016, microsoft’s racist chatbot revealed the dangers of online
conversation - ieee spectrum,” https://spectrum.ieee.org/in-2016-
microsofts-racist- chatbot-revealed-the- dangers-of-online- conversation,
(Accessed on 06/29/2023).
[79] H. M. Alhaidry, B. Fatani, J. O. Alrayes, A. M. Almana, N. K. Alfhaed,
H. Alhaidry, J. Alrayes, A. Almana, and N. K. Alfhaed Sr, “Chatgpt in
dentistry: A comprehensive review,” Cureus, vol. 15, no. 4, 2023.
[80] T. T. Nguyen, N. Larrivée, A. Lee, O. Bilaniuk, and R. Durand, “Use
of artificial intelligence in dentistry: current clinical trends and research
advances,” J Can Dent Assoc, vol. 87, no. l7, pp. 1488–2159, 2021.
[81] “The attribution problem with generative ai - hacking semantics,” https:
//hackingsemantics.xyz/2022/attribution/, (Accessed on 07/17/2023).
[82] S. Harrer, “Attention is not all you need: the complicated case of ethically
using large language models in healthcare and medicine,” EBioMedicine,
vol. 90, 2023.
[83] B. Perrigo, “Openai used kenyan workers on less than 2 dollars per hour
to make chatgpt less toxic. time,” 2023.
[84] N. Martinez-Martin, Z. Luo, A. Kaushal, E. Adeli, A. Haque, S. S. Kelly,
S. Wieten, M. K. Cho, D. Magnus, L. Fei-Fei et al., “Ethical issues in using
ambient intelligence in health-care settings,” The lancet digital health,
vol. 3, no. 2, pp. e115–e123, 2021.
[85] A. L. Allen, “Coercing privacy,” Wm. & Mary L. Rev., vol. 40, p. 723,
1998.
VOLUME X, 2020 27
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
[86] J. Deng and Y. Lin, “The benefits and challenges of chatgpt: An overview,”
Frontiers in Computing and Intelligent Systems, vol. 2, no. 2, pp. 81–83,
2022.
[87] S. Goldman, “Stable diffusion ai art lawsuit, plus caution
from openai, deepmind | the ai beat,” 2023. [Online]. Avail-
able: https://venturebeat.com/ai/stable-diffusion-lawsuit- plus-words-of-
caution-from- openai-deepmind-the-ai-beat/
[88] P. J. B. Aaron J Snoswell, “A galaxy of deep science fakes: The problems
with galactica ai,” 2022. [Online]. Available: https://www.siliconrepublic.
com/machines/galactica-ai- meta-fake-science- misinformation
SIVA SAI is a research scholar in the EEE depart-
ment at Birla Institute of Technology and Science
(Pilani). Previously, he has completed his B.E in
Computer Science and M.Sc (Hons) in Economics
with BITS Pilani. His research interests include
applications of blockchain and machine learning
for healthcare, natural language processing, com-
puter vision, and connected vehicles. He worked
on multiple research problems, including WiFi
CSI activity recognition, multimodal hate speech
detection, multilingual offensive/fake speech identification, NLP technolo-
gies for lower-resource languages, and deep learning for time series analysis.
His past publications were accepted by prestigeous journals and conferences
like EACL, AAAI, FIRE, EMNLP, IEEE IoT journal, IEEE TITS, and
Neural Networks(Elsevier).
AANCHAL GAUR is a final-year student in the
ECE department at Maharaja Agrasen Institute of
Technology (New Delhi). She is a research intern
in medsupervision at the Birla Institute of Tech-
nology and Science (Pilani). Her research interests
include machine learning and artificial intelligence
applications in healthcare, agriculture, and com-
puter vision. Throughout her academic journey,
she has actively pursued challenging projects in
machine learning. She is dedicated to exploring
the applications of cutting-edge technology in healthcare.
REVANT SAI is currently a pre-final year stu-
dent pursuing a Bachelor of Engineering degree
in Computer Science at Birla Institute of Technol-
ogy and Science, Pilani. With a keen interest in
the ever-evolving fields of Artificial Intelligence
and Machine Learning, he has actively pursued
coursework and practical projects to advance their
knowledge and skills. Throughout his academic
journey, Revant Sai has undertaken challenging
projects in areas including Database Systems,
Object-Oriented Programming (OOP), and Data Structures. These experi-
ences have not only equipped him with a strong foundation in computer
science fundamentals but have also cultivated a problem-solving mindset
crucial in the world of AI and ML. Revant’s research interests revolve around
the convergence of data-driven decision-making and machine intelligence.
He is dedicated to exploring innovative solutions to intricate problems
and harnessing the capabilities of AI and ML to better comprehend the
complexities of our world. With a firm belief in the transformative potential
of technology, Revant Sai is committed to contributing to the advancement
of these fields for the benefit of society.
28 VOLUME X, 2020
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
VINAY CHAMOLA (Fellow, IET) received the
B.E. degree in electrical and electronics engineer-
ing and master’s degree in communication engi-
neering from the Birla Institute of Technology and
Science, Pilani, India, in 2010 and 2013, respec-
tively. He received his Ph.D. degree in electrical
and computer engineering from the National Uni-
versity of Singapore, Singapore, in 2016. In 2015,
he was a Visiting Researcher with the Autonomous
Networks Research Group (ANRG), University of
Southern California, Los Angeles, CA, USA. He also worked as a post-
doctoral research fellow at the National University of Singapore, Singapore.
He is currently an Associate Professor with the Department of Electrical and
Electronics Engineering, BITS-Pilani, Pilani, where he heads the Internet of
Things Research Group / Lab. His research interests include IoT Security,
Blockchain, UAVs, VANETs, 5G, and Healthcare. He serves as an Area
Editor for the Ad Hoc Networks Journal, Elsevier and the IEEE Internet of
Things Magazine. He also serves as an Associate Editor in the IEEE Transac-
tions on Intelligent Transportation Systems, IEEE Networking Letters, IEEE
Consumer electronics magazine, IET Quantum Communications, IET Net-
works, and several other journals. He serves as co-chair of various reputed
workshops like IEEE Globecom Workshop 2021, IEEE INFOCOM 2022
workshop, IEEE ANTS 2021, and IEEE ICIAfS 2021, to name a few. He is
listed in the World’s Top 2% Scientists identified by Stanford University. He
is co-founder and President of a healthcare startup Medsupervision Pvt. Ltd.
He is a senior member of the IEEE.
MOHSEN GUIZANI (S’85–M’89–SM’99–F’09)
received the B.S. (with distinction), M.S. and
Ph.D. degrees in Electrical and Computer engi-
neering from Syracuse University, Syracuse, NY,
USA. He is currently a Professor at the Ma-
chine Learning Department, Mohamed Bin Zayed
University of Artificial Intelligence (MBZUAI),
Abu Dhabi. Previously, he worked in different
institutions: Qatar University, University of Idaho,
Western Michigan University, University of West
Florida, University of Missouri-Kansas City, University of Colorado-
Boulder, and Syracuse University. His research interests include wireless
communications and mobile computing, applied machine learning, cloud
computing, security and its application to healthcare systems. He was
elevated to the IEEE Fellow in 2009. He was listed as a Clarivate Analytics
Highly Cited Researcher in Computer Science in 2019 and 2020. Dr. Guizani
has won several research awards including the “2015 IEEE Communications
Society Best Survey Paper Award” as well 4 Best Paper Awards from ICC
and Globecom Conferences. He is the author of nine books and more than
800 publications. He is also the recipient of the 2017 IEEE Communications
Society Wireless Technical Committee (WTC) Recognition Award, the
2018 AdHoc Technical Committee Recognition Award, and the 2019 IEEE
Communications and Information Security Technical Recognition (CISTC)
Award. He served as the Editor-in-Chief of IEEE Network and is currently
serves on the Editorial Boards of many IEEE journals/Transactions. He
was the Chair of the IEEE Communications Society Wireless Technical
Committee and the Chair of the TAOS Technical Committee. He served as
the IEEE Computer Society Distinguished Speaker and is currently the IEEE
ComSoc Distinguished Lecturer.
JOEL J. P. C. RODRIGUES ( (Fellow, IEEE) is
a Leader of the Center for Intelligence at Fecomér-
cio/CE, Brazil, and Full Professor at COPELABS,
Lusófona University, Lisbon, Portugal. He is an
Highly Cited Researcher (Clarivate), N. 1 of the
top scientists in computer science in Brazil (Re-
search.com), the Leader of the Next Generation
Networks and Applications (NetGNA) research
group (CNPq), Member Representative of the
IEEE Communications Society on the IEEE Bio-
metrics Council, and the President of the scientific council at ParkUrbis
– Covilhã Science and Technology Park. He has authored or coauthored
about 1150 papers in refereed international journals and conferences, three
books, two patents, and one ITU-T Recommendation. He was the Director
for Conference Development - IEEE ComSoc Board of Governors, an IEEE
Distinguished Lecturer, Technical Activities Committee Chair of the IEEE
ComSoc Latin America Region Board, a Past-Chair of the IEEE ComSoc
Technical Committee (TC) on eHealth and the TC on Communications
Software, a Steering Committee member of the IEEE Life Sciences Tech-
nical Community and Publications co-Chair. He is a member of the Internet
Society, a senior member ACM, and Fellow of AAIA. He has been general
chair and TPC Chair of many international conferences, including IEEE
ICC, IEEE GLOBECOM, IEEE HEALTHCOM, and IEEE LatinCom. He
had been awarded several Outstanding Leadership and Outstanding Service
Awards by IEEE Communications Society and several best papers awards.
He is the Editor-in-Chief of the International Journal of E-Health and
Medical Communications and a editorial board member of several high
reputed journals (mainly, from IEEE).)
VOLUME X, 2020 29
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2024.3367715
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
