Content uploaded by Nalan Karunanayake
Author content
All content in this area was uploaded by Nalan Karunanayake on Apr 10, 2025
Content may be subject to copyright.
Next-generation agentic AI for transforming healthcare
Nalan Karunanayake
Department of Radiology, Memorial Sloan Kettering Cancer Center, New York, NY 10065, USA
ARTICLE INFO
Keywords
Articial intelligence
Agentic AI
AI agents
Healthcare
Personalized medicine
ABSTRACT
Articial Intelligence (AI) is transforming the healthcare landscape, yet many current applications remain
narrowly task-specic, constrained by data complexity and inherent biases. This paper explores the emergence of
next generation "agentic AI" systems, characterized by advanced autonomy, adaptability, scalability, and prob-
abilistic reasoning, which address critical challenges in medical management. These systems enhance various
aspects of healthcare, including diagnostics, clinical decision support, treatment planning, patient monitoring,
administrative operations, drug discovery, and robotic-assisted surgery. Powered by multimodal AI, agentic
systems integrate diverse data sources, iteratively rene outputs, and leverage vast knowledge bases to deliver
context-aware, patient-centric care with heightened precision and reduced error rates. These advancements
promise to enhance patient outcomes, optimize clinical workows, and expand the reach of AI-driven solutions.
However, their deployment introduces ethical, privacy, and regulatory challenges, emphasizing the need for
robust governance frameworks and interdisciplinary collaboration. Agentic AI has the potential to redene
healthcare, driving personalized, efcient, and scalable services while extending its impact beyond clinical
settings to global public health initiatives. By addressing disparities and enhancing care delivery in resource-
limited environments, this technology could signicantly advance equitable healthcare. Realizing the full po-
tential of agentic AI will require sustained research, innovation, and cross-disciplinary partnerships to ensure its
responsible and transformative integration into healthcare systems worldwide.
1. Introduction
The rapid advancement of articial intelligence (AI) is reshaping
global healthcare, positioning digital technologies as integral to modern
medical systems. We are entering an ’agentic era,’ characterized by AI
agent systems capable of autonomous functionality, advanced
reasoning, and dynamic human–AI interactions. These advancements
are driven by the integration of Multimodal Large Language Models
(MLLMs) with other sophisticated AI systems, as demonstrated by AI
assistants like OpenAI’s ChatGPT (based on GPT-4), Google’s Gemini
(built on the Gemini 1.5 model), and Microsoft’s Copilot (which utilizes
GPT-4 and other AI models).
AI’s relationship with healthcare spans decades, from personalized
care to drug discovery,
1,2
beginning with pioneering rule-based systems
in the 1970s, such as MYCIN,
3
INTERNIST-1, and QMR,
4
DXplain,
5
which addressed diagnostic challenges but struggled with the growing
complexity of medical knowledge. By the late 1990s, the advent of larger
healthcare datasets, improved computational capacity, and advanced
machine learning (ML) algorithms marked a shift to data-driven ap-
proaches.
6
This collaboration between ML researchers and medical
professionals fostered clinically relevant automation.
7
The rise of deep
learning (DL) in the 2000s further accelerated progress, particularly in
medical imaging, where convolutional neural networks (CNNs) enabled
accurate anomaly detection, image segmentation, and classication.
8–12
These innovations signicantly improved diagnostic precision and
prognostic predictions.
Despite these breakthroughs, AI integration in healthcare faces
challenges beyond medical data complexity, including interoperability
with clinical workows, regulatory constraints, and integration with
medical devices.
13,14
Healthcare applications demand accuracy, reli-
ability, and context-aware reasoning to ensure patient safety. While
current AI systems excels in large-scale pattern recognition,
15
they still
face challenges in higher-level reasoning and decision-making.
16
To
address these complexities, AI agents designed for decision support can
offer a solution, provided they are integrated with evidence-based
strategies to ensure reliability and transparency in healthcare
applications.
AI agents extend beyond traditional rule-based systems by dynami-
cally optimizing workows, adapting to tasks with minimal human
intervention, and integrating with specialized AI models. Depending on
E-mail address: karunan@mskcc.org.
Contents lists available at ScienceDirect
Informatics and Health
journal homepage: www.keaipublishing.com/en/journals/informatics-and-health/
https://doi.org/10.1016/j.infoh.2025.03.001
Received 18 January 2025; Received in revised form 12 March 2025; Accepted 19 March 2025
Informatics and Health 2 (2025) 73–83
Available online 7 April 2025
2949-9534/© 2025 The Author(s). Publishing services by Elsevier B.V. on behalf of KeAi Communications Co. Ltd. This is an open access article under the CC BY
license ( http://creativecommons.org/licenses/by/4.0/ ).
their architecture, these agents can operate independently or collaborate
within multi-agent frameworks to address complex workows.
17
Their
applications span healthcare domains, from diagnostics to drug dis-
covery, demonstrating their transformative potential in advancing
global healthcare (detailed in Section 2).
1.1. Traditional AI agents vs. agentic AI systems
Traditional AI agents, as classied in,
18,19
fall into six categories:
1. Simple Reex Agents: Operate on condition–action rules without
memory.
2. Model-Based Reex Agents: Extend reex agents with an internal
model that updates based on past percepts.
3. Goal-Based Agents: Plan and execute actions to achieve specied
objectives.
4. Utility-Based Agents: Use utility functions to choose actions that
maximize expected utility, managing conicting goals.
5. Learning Agents: Continuously improve strategies based on new
experiences.
6. Problem-Solving Agents: Employ search algorithms to achieve
desired outcomes.
Traditional AI agents typically utilize rule-based systems, heuristic-
driven ML, and, in some cases, reinforcement learning (RL) for deci-
sion optimization, each optimized for predened objectives. In health-
care, these approaches have been applied to clinical decision support,
diagnostics, and treatment planning, but they often struggle with
adaptability in dynamic healthcare settings. In contrast, agentic AI ad-
vances these paradigms by incorporating adaptive architectures that
enhance decision-making autonomy and exibility. By enabling AI
agents to generalize across diverse healthcare tasks—such as personal-
ized treatment optimization and autonomous medical imaging ana-
lysis—these advancements represent progress toward more versatile AI
systems, aligning with long-term goals in Articial General Intelligence
(AGI) research.
Table 1 summarizes the differences between traditional and agentic
AI in healthcare. An illustrative application is the AI healthcare agent
(Fig. 1), which serves as a hub for processing data from medical data-
bases, electronic health records (EHRs), medical scans, and neural net-
works. By generating prompts and offering insights, these agents support
high-stakes decision-making in clinical settings, showcasing the trans-
formative potential of agentic AI in healthcare and beyond.
1.2. Technical overview of agentic AI
Agentic AI systems often integrate pretrained DL encoders for pro-
cessing multimodal data, such as images and text, though alternative
approaches exist. These encoded representations are processed by a
central LLM, which functions as the reasoning and decision-making core
of the agent. Commonly employed open-source LLMs in this domain
include LLaMA,
20
Falcon
21
and Vicuna,
22
which have been pretrained
on large-scale datasets. The openness of these models allows researchers
to develop their own AI agents with greater transparency and a deeper
understanding of their internal reasoning and decision-making pro-
cesses. In contrast, proprietary LLMs such as the ChatGPT models and
Gemini, while often delivering highly optimized performance and
broader support, offer minimal transparency regarding their internal
decision-making procedure.
Key mechanisms leveraged by these central LLMs include advanced
reasoning techniques, such as chain-of-thought (CoT),
23
reasoning and
acting (ReAct),
24
tree of thought (ToT).
25
These techniques enhance
reasoning and decision-making through structured decomposition,
logical inference, and contextual understanding.
A critical advantage of agentic AI lies in its modularity and adapt-
ability, allowing seamless integration of LLMs, DL models, and other
specialized AI components. This modular architecture supports scal-
ability, enhances exibility in adapting to evolving tasks, and enables
deployment across diverse application domains, including autonomous
systems, scientic research, and complex decision-making frameworks.
2. Signicance of AI agents in healthcare
Recent advances in AI-related medical technologies have fueled
unprecedented growth in healthcare data. Between 2011 and 2018,
medical image datasets grew by three- to ten-fold, with annual increases
of 21–32 % across various imaging modalities.
26
The proliferation of
mobile applications, cloud computing, wearable devices, and big-data
analytics has further expanded data sources, enabling personalized
medicine and real-time health monitoring.
27
Despite these advance-
ments, the healthcare sector faces a projected shortfall of 18 million
workers by 2030, particularly in low-income regions.
28,29
AI agents can
help mitigate this gap by integrating advanced ML, DL, natural language
Table 1
Key Differences Between Traditional AI Agents and Agentic AI in healthcare.
Aspect Traditional AI agent Agentic AI
Reasoning
Approach
Use domain specic algorithms
(e.g., RL, nite state machines).
Example: A rule-based chatbot
that answers predened
medical FAQs.
Natural language-based
reasoning, primarily leverage
LLMs.
Example: An AI assistant that
engages in open-ended medical
Q&A and explains conditions
based on patient history.
Domain
Flexibility
Optimized for a specic task or
environment.
Example: Optimized for
specic tasks such as structured
report generation or rule-based
classication.
Adaptable to multiple tasks
using zero/few-shot learning or
dynamic prompt engineering.
Example: Can summarize
radiology reports and also
assist in diagnostic support.
Decision
Structure
Rule-based, with a predened
set of actions and goals.
Example: An AI system that
follows xed guidelines for
triaging emergency cases.
Capable of self-directed sub-
goals and iterative planning (e.
g., language-based reasoning
models).
Example: An AI agent that
dynamically adjusts triage
decisions based on patient data
and evolving symptoms.
Data &
Training
Domain-specic datasets.
Example: AI trained only on
structured EHR data.
Utilizes large-scale pretraining
over vast corpora; can
generalize or specialize with
minimal updates (few-shot).
Example: AI trained on diverse
medical texts, adapting quickly
to new diseases and guidelines.
Tool
Integration
Limited.
Example: A standalone AI tool
for detecting anomalies in CT
scans.
Dynamically calls external
APIs, knowledge bases, ML, DL,
MLLMs.
Example: An AI system that
integrates with hospital
systems, lab results, and
medical literature to provide
comprehensive diagnostic
insights.
Explanations Transparent in symbolic
systems; moderate
interpretability in RL.
Example: A decision tree model
explaining why a drug is
recommended.
Largely blackbox, though
techniques for interpretability
are emerging.
Example: Agentic AI system
suggesting treatments based on
DL patterns, with efforts to
explain its reasoning using
attention maps or
counterfactual analysis.
Primary Use
Cases
Robotics, rule-based decision-
making, industrial control
systems.
Example: A robotic system
following predened surgical
procedures.
Unlimited
Example: AI-driven virtual
assistants for patient
monitoring, adaptive treatment
planning, personalized
medicine, and robotic surgery
and many more.
N. Karunanayake
Informatics and Health 2 (2025) 73–83
74
processing (NLP), and computer vision technologies, automating
administrative tasks, enhancing diagnostics, and improving workow
efciency.
2.1. Features of agentic AI in healthcare
Fig. 2 illustrates the key features that make agentic AI technologies
well-suited for healthcare environments.
Autonomy: AI agents function independently, making decisions
based on predened goals and real-time data inputs. In radiology, for
instance, an AI agent can autonomously analyze scans, select diagnostic
algorithms, and generate preliminary reports. In multilingual settings,
the agent can detect the need for translation and invoke appropriate
tools, reducing manual intervention and expediting workows.
Adaptability: Unlike traditional AI models that are typically opti-
mized for specic tasks, agentic AI dynamically adjusts to new data and
evolving clinical needs. For example, an AI agent trained on X-ray
analysis can ne-tune itself to process MRI or CT scans, ensuring
continued relevance as imaging modalities evolve.
Scalability: By leveraging cloud infrastructures and federated
learning, agentic AI can handle vast and heterogeneous data in real time.
This capability is critical for applications like telemedicine, where large-
scale data analysis must be performed without compromising speed or
accuracy.
Probabilistic Decision-Making: AI agents use iterative reasoning,
continuously updating their predictions based on new data, contextual
knowledge, and feedback loops. For instance, an AI agent might initially
diagnose pneumonia but revise it to tuberculosis after incorporating
Fig. 1. Healthcare AI agent: Integrating multimodal data for collaborative clinical decision making.
Fig. 2. Key agentic AI features in healthcare.
N. Karunanayake
Informatics and Health 2 (2025) 73–83
75
epidemiological data and lab results.
These features collectively enable agentic AI to provide robust,
efcient, and context-aware solutions across diverse healthcare settings.
2.2. Core functional domains of AI agents in healthcare
As illustrated in Fig. 3, healthcare-focused AI agents can be catego-
rized into several core functional domains, including diagnosis, clinical
decision support, treatment and patient care, patient engagement and
monitoring, operations and administration, drug discovery and
research, and robot-assisted surgery. The following subsections describe
how AI and AI agents are applied within each of these domains in
healthcare.
2.2.1. AI agents in diagnosis
Modern diagnostic workows rely heavily on AI for medical image
analysis and predictive analytics. The growing volume of imaging data
and complexity of EHRs can overwhelm clinicians, but AI helps alleviate
these burdens,
30,31
reducing cognitive load and enhancing patient
safety.
32,33
AI integration with medical imaging is transforming diagnostics by
improving accuracy. In radiology, AI enhances diagnostic accuracy
through anomaly detection,
34,35
lesion segmentation,
36–38
classica-
tion.
39
DL technologies like CNNs and vision transformers (ViTs) opti-
mize workows within PACS for modalities such as CT, MRI, X-ray, and
pathological imaging, while maintaining patient data condentiality.
40
Additionally, AI uses MLLMs to integrate radiological and clinical data,
supporting diagnoses and automating report generation.
41–44
Beyond radiology, AI excels in analyzing pathology images,
enhancing cancer detection and microscopic diagnoses.
45
A recent study
46
explores the role of text-only AI agents in assisting
radiologists with brain MRI differential diagnosis. A comparison be-
tween LLM-assisted agents and conventional internet search workows
showed that AI-assisted diagnoses had higher accuracy (61.4 % vs.
46.5 %), though there was no signicant difference in interpretation
time or condence levels. Integrating medical images into a MLLM could
potentially eliminate the need for manual descriptions of imaging
ndings, further enhancing diagnostic efciency. For accurate diagnosis
image analysis is a key criterion. Med-Flamingo
47
is a multimodal
few-shot AI agent designed for medical diagnosis, using 2D images and
text to assist clinicians. It outperforms existing models in diagnostic
accuracy by up to 20 %. Beyond 2D imaging, M3D-LaMed
48
is an MLLM
designed for 3D medical image analysis, integrating text-based data to
enhance diagnostic tasks. By leveraging a pre-trained 3D vision encoder
and an efcient 3D spatial pooling perceiver, M3D-LaMed improves
diagnosis for complex 3D medical images like CT and MRI scans, rep-
resenting a signicant advancement in AI-assisted medical diagnosis.
The integration of AI agents into imaging workows has the poten-
tial to optimize operations, enhance diagnostic accuracy, and streamline
processes. In radiology, AI agents can interface with PACS to automate
quality assurance, manage data transfers, and execute DL and ML al-
gorithms for image analysis. These agents ag abnormalities for review
while maintaining patient privacy through de-identication and can
operate in the background, continuously analyzing imaging data,
generating preliminary reports, and proposing diagnoses using MLLMs.
By combining computational efciency with radiologists’ expertise, AI
agents improve accuracy, reduce uncertainty, and save time and
resources.
Fig. 3. Key applications of AI agent in healthcare industry.
N. Karunanayake
Informatics and Health 2 (2025) 73–83
76
Beyond diagnostics, AI agents advance predictive analytics and
personalized care by analyzing EHRs, imaging, genomics, and wearable
device data. They identify subtle disease indicators, such as early signs
of diabetes,
49,50
cardiovascular conditions,
51
or cancer,
52
enabling early
detection, optimized resource allocation, and improved patient out-
comes. AI agents also enhance prognosis by forecasting disease trajec-
tories, modeling progression, and tailoring treatment plans.
53–56
These
insights help physicians anticipate complications and adjust therapies
proactively.
For effective integration, AI agents must embed seamlessly into
clinical workows and patient environments. In clinical settings, they
provide real-time data and actionable insights, while on the patient side,
wearables equipped with AI monitor health metrics and offer person-
alized recommendations, bridging clinical care and daily life. Addi-
tionally, AI agents serve as educational tools, creating interactive
learning modules, quizzing students on realistic cases
57,58
and sup-
porting clinical decision-making, resource management, and healthcare
professional training.
2.2.2. AI agents in clinical decision support systems
AI can be seamlessly integrated into clinical decision support systems
(CDSS) to enhance patient care.
59,60,54
By analyzing EHRs, clinical data,
genomic data, behavioral data, and administrative data,
14
AI systems
summarize key ndings, establish connections using LLMs, access
medical databases, generate reminders, and facilitate collaboration.
These capabilities improve diagnoses, reduce misdiagnoses, and ease
clinicians’ workloads through automation and task prioritization.
The incorporation of healthcare AI agents into CDSS accessible via
mobile devices as standalone tools or part of a multi-agent system rep-
resents a major advancement. Unlike traditional CDSS, where tasks such
as information retrieval and summarization are performed manually or
in separate steps, AI agents can execute these processes cohesively and
adaptively. They rene outputs by evaluating individual AI module re-
sults,
61
delivering more precise and reliable clinical insights.
Vision-Language Models (VLMs) integrate LLMs with medical im-
aging, aiding clinicians in interpreting images and generating structured
reports.
62
A study explored the use of AI-driven VLMs in radiology,
identifying key applications such as draft report generation, augmented
report review, visual search and querying, and patient imaging history
highlights. The ndings suggest that AI-assisted report generation and
visual search functionalities can improve workow efciency, reduce
cognitive burden, and enhance diagnostic accuracy. However, the study
highlights the need for explainability, seamless workow integration,
and clinician oversight to ensure AI-generated insights are clinically
reliable and interpretable. Another VLM, LLaVA-Med, an AI agent, was
trained using a large-scale biomedical dataset from PubMed Central and
GPT-4-based instruction tuning to support biomedical image interpre-
tation, clinical reporting, and medical visual question answering
(VQA).
63
The study demonstrated that LLaVA-Med outperforms previ-
ous state-of-the-art models on standard biomedical VQA datasets,
proving its efcacy in assisting radiologists and clinicians.
As the rst AI agent in medical imaging that dynamically plans,
executes, and adapts using external computational tools, VoxelPrompt
61
breaks free from rigid, single-task models. It outperforms specialized
segmentation models on 13/17 anatomical structures, surpasses con-
ventional vision models on 23 different brain regions, and achieves 89 %
accuracy in pathology characterization—matching expert classiers
while handling a broader range of tasks that helps clinical decision
making. Another practical example the study
64
surpasses traditional AI
by acting as an intelligent agent named LLMSeg that fuses textual clin-
ical data with imaging, enabling precise 3D contouring. This approach
achieves superior accuracy with signicantly less data, maintaining
robust performance even data-insufcient settings where conventional
AI struggles. By leveraging LLMs, AI agents enhance efciency and
adaptability in clinical settings, improving automated report generation,
decision support, and disease characterization.
Despite signicant advancements in agentic AI for CDSS, many AI-
driven decision support systems remain at the research level and have
yet to achieve widespread clinical adoption. A study
65
address this
challenge by proposing a systematic AI support framework that con-
siders key dimensions such as disease, data, technology, user groups,
validation, decision-making, and maturity, emphasizing a structured
approach to improving the real-world integration of AI in healthcare.
As AI agents continue to evolve, human-centered AI design and
regulatory considerations will be critical for ensuring trust, usability,
and compliance in CDSS. The integration of AI-powered decision sup-
port systems in radiology and other medical domains has the potential to
revolutionize diagnostics, reduce clinician workload, and improve pa-
tient outcomes.
2.2.3. AI agents in treatment and patient care
AI has transformative potential in treatment and patient care
through advanced functionalities. DL and ML models enable personal-
ized medicine by analyzing genomics and clinical data to create tailored
treatment plans,
66
simulate drug responses,
67
and optimize chemo-
therapy and radiation therapy protocols for maximum efcacy.
68
DL
also contributes to therapeutic optimization.
69
Radiomics combined
with AI further enhances treatment planning,
70
while AI embedded in
medical devices ensures precision monitoring, reduces errors, and en-
ables remote and in-hospital patient surveillance. These systems alert
caregivers to critical changes and assist rst responders with real-time,
context-specic guidance during emergencies.
71,72
The study
73
presents AgentClinic, an AI agent that simulates
doctor-patient interactions, autonomously collecting patient informa-
tion, making diagnostic decisions, and recommending treatments.
AgentClinic actively engages with patients through dialogue, requests
medical tests, and adapts its questioning strategies in real time,
mimicking human clinical decision-making. The study highlights how
these AI-driven agentic interactions inuence patient engagement and
monitoring, particularly in shaping patient compliance, trust in AI rec-
ommendations, and follow-up willingness.
Currently, AI technologies for treatment operate as standalone al-
gorithms tailored to specic datasets and tasks. Transitioning to AI
agents allows for seamless integration with existing models, databases,
and other AI modules, improving efciency. In patient care, traditional
semi-autonomous AI modules often result in fragmented workows and
errors. AI agents overcome these challenges by integrating smoothly
into healthcare systems. They analyze genomic proles and patient
histories to recommend personalized treatments, validate these with
extensive databases, and optimize therapy plans by simulating compli-
cations and alerting clinicians proactively.
2.2.4. AI agents in patient engagement and monitoring
AI plays a transformative role in patient engagement and monitoring,
leveraging advanced technologies to enhance care. In remote moni-
toring, AI analyzes data from wearables and home-based devices, such
as cameras and sensors, to track vital signs and detect abnormalities in
real time.
74
Virtual health assistants and chatbots provide symptom
checks, automated triage, and scheduling guidance, reducing the
workload of healthcare professionals.
75,76
However, many chatbots are
limited by the size and heterogeneity of their training datasets, which
restrict adaptability. On telehealth platforms, AI improves video
consultation scheduling and performs real-time patient data analysis,
enabling efcient and personalized virtual care.
77,78
Recent study
79
presents Agent Hospital, an AI agent system that
enhances patient engagement by simulating interactive doctor-patient
interactions in a virtual hospital. LLM agents autonomously manage
triage, consultation, and follow-ups, engaging patient agents through
dynamic dialogues, medical examinations, and feedback loops. The
study highlights how AI agents track patient responses, adapt decisions
based on evolving symptoms, and inuence follow-up adherence,
demonstrating their potential in proactive healthcare monitoring.
N. Karunanayake
Informatics and Health 2 (2025) 73–83
77
To overcome the limitations of traditional AI systems, AI agents
seamlessly integrate into healthcare systems and devices. Paired with
wearables and home monitoring tools, AI agents continuously analyze
health data, alert caregivers when needed, and provide patients with
emergency instructions. Virtual health agents on mobile platforms guide
users through symptom assessments, triage, and scheduling, while
aligning appointments with personal calendars for a smooth healthcare
experience. On telehealth platforms, AI agents optimize schedules,
analyze patient data during video calls, and support clinicians with
advanced features like real-time language translation and sign language
interpretation using DL and NLP. These capabilities ensure accessibility,
efciency, and precision in patient engagement and monitoring,
bridging cutting-edge technology with personalized care.
AI conversational agents have emerged as valuable tools for mental
health monitoring and patient engagement. A notable example is Woe-
bot, an AI-powered chatbot that dynamically enhances user engagement
by actively analyzing emotional states through NLP and delivering
context-aware, empathic responses in real time.
80
Another study eval-
uates the use of LLM agent in mental health support,
81
revealing biases
in empathy and response quality across different demographic groups.
While LLMs like GPT-4 can enhance patient engagement by encouraging
behavior change, the study highlights the need for bias mitigation
strategies to ensure equitable and ethical deployment in clinical settings.
The AI agent tailors therapeutic content by continuously adapting its
dialogue based on user inputs, fostering an interactive and personalized
support system. Through proactive daily check-ins and automated mood
tracking, monitors user sentiment, detects uctuations in emotional
well-being, and delivers targeted interventions to promote self-
reection, emotional awareness, and sustained participation. This un-
derscores the potential of AI agents to serve as engaging, scalable, and
readily available platforms for continuous patient engagement and
monitoring, augmenting and complementing traditional therapeutic
methods.
2.2.5. AI agents in healthcare operations and administration
DL and ML technologies enhances healthcare operations through
advanced data analysis, task automation, and predictive modeling.
82–85
In workow optimization, AI streamlines triage in emergency de-
partments, automates patient ow and bed allocation,
86
facilitates lan-
guage translation and disability support,
87
and provides virtual
assistance for training and management of healthcare staff.
88
In billing
and coding, AI automates claim processing, veries insurance details,
89
and detects fraudulent activities.
90
Predictive analytics further optimize
scheduling and resource management by improving staff rostering,
operating room planning, and resource allocation.
91
AI agents enhance healthcare operations by integrating with hospital
systems, optimizing workows, and reducing administrative burdens. In
emergency rooms, they monitor real-time data to prioritize patients,
manage bed assignments, and coordinate transfers. Leveraging LLMs, AI
agents support communication through language translation and
disability accommodations. They also assist healthcare staff with
training, updated protocols, and best practices. In nancial operations,
AI agents automate billing, claims verication, and fraud detection,
reducing administrative burdens. For scheduling, they analyze stafng
needs, optimize rosters, plan surgeries, and forecast resources, dynam-
ically adjusting to maintain operational efciency and ensure high-
quality patient care. A practical example of AI agent integration in
healthcare operations is NYUTron,
92
the rst clinically deployed AI
agent used in real-world hospital settings. Embedded within NYU Lan-
gone Health’s EHR system, NYUTron assists physicians and adminis-
trators by predicting hospital readmission risk, length of stay, in-hospital
mortality, and insurance claim denials, showcasing how AI agents can
enhance decision-making and streamline healthcare operations.
Another practical example of AI agent integration in healthcare opera-
tions is GPT4DFCI,
93
an institute-wide AI deployment at Dana-Farber
Cancer Institute. Designed as a secure, HIPAA-compliant AI system,
GPT4DFCI supports research, clinical documentation, and administra-
tive workows while addressing challenges in data privacy, regulatory
compliance, and ethical governance. This initiative demonstrates how
AI agents can be safely integrated into healthcare institutions to enhance
efciency while maintaining strict oversight and responsible AI use.
2.2.6. AI agents in drug discovery and research
Advancements in AI are revolutionizing drug discovery by enhancing
efciency and precision. In drug candidate screening, AI-powered high-
throughput screening rapidly analyzes vast compound libraries, while
molecular modeling predicts interactions with biological targets, nar-
rowing down potential candidates.
94
These models require extensive
datasets and expert domain knowledge for effective training and eval-
uation.
95
In clinical trial optimization, AI analyzes patient data to
improve recruitment, designs adaptive protocols based on emerging
data, and renes parameters to boost success rates.
96,97
Additionally, AI
processes large-scale genomic data to identify genetic markers, enabling
precision therapeutics tailored to individual proles.
98,99
These in-
novations shorten discovery timelines, reduce costs, and advance
personalized medicine.
AI agents further accelerate drug discovery by automating complex
workows and reducing errors. In candidate screening, agents conduct
high-throughput experiments, run molecular simulations, and analyze
results to identify promising compounds. For clinical trials, they
leverage EHRs and genomic databases to recruit suitable participants,
streamline processes, and adapt protocols in real time. In genomics, AI
agents analyze sequencing data, detect disease-linked mutations, and
propose personalized therapies. Integrated with lab instruments, data
platforms, and clinical systems, AI agents automate decision-making,
enhance accuracy, and expedite the development of safe, effective
treatments.
In a recent study,
100
a multi-agent AI system named DrugAgent was
designed to automate ML programming for drug discovery. By inte-
grating domain-specic knowledge with AI-driven model selection, it
enhances key tasks such as ADMET prediction, drug-target interaction
analysis, and molecular optimization. Unlike traditional non-agentic
methods, which require manual coding and expert intervention, Drug-
Agent autonomously generates, tests, and renes ML models, signi-
cantly improving efciency and reducing errors. The model achieved an
F1 score of 0.92 in drug absorption prediction, demonstrating its ability
to optimize predictive models for pharmaceutical research.
Beyond ML-driven drug discovery, agentic AI systems are trans-
forming biomedical research by autonomously generating, rening, and
validating scientic hypotheses, thereby saving signicant time. The AI
co-scientist
101
exemplies this shift by leveraging self-improving, mul-
ti-agent reasoning to advance drug repurposing, novel drug target dis-
covery, and antimicrobial resistance research. The AI co-scientist
identied repurposed drug candidates for acute myeloid leukemia
(AML), for novel target discovery, it proposed epigenetic regulators for
liver brosis, rening hypotheses through expert feedback and experi-
mental validation. Additionally, it independently hypothesized a bac-
terial gene transfer mechanism, aligning with unpublished
microbiological ndings, showcasing its autonomous reasoning and
discovery capabilities. By integrating human in the loop (HITL) collab-
oration, the AI co-scientist ensures expert alignment while continuously
evolving through a self-improving feedback loop, marking a new era of
AI-empowered scientic discovery in drug development. These exam-
ples highlights how AI agents can streamline drug development, reduce
human workload, and accelerate pharmaceutical innovation.
2.2.7. AI agents in robot-assisted surgery
AI-guided robot-assisted surgery marks a new era in medicine,
enhancing precision, accuracy, and safety in complex procedures.
102,103
By integrating advanced imaging, sensor data, ML, and robotics, AI
systems enable surgical robots to perform tasks with exceptional preci-
sion. AI-driven navigation and instrument guidance allow real-time
N. Karunanayake
Informatics and Health 2 (2025) 73–83
78
interpretation of patient anatomy, adaptation to subtle movements, and
optimal instrument positioning, reducing invasiveness and minimizing
risks. These advancements support surgeons in planning, executing, and
rening interventions, improving outcomes and expanding minimally
invasive surgery possibilities.
AI agents embedded in robotic systems orchestrate these capabilities
by analyzing data from endoscopic cameras, sensors, and preoperative
imaging to chart precise surgical paths and guide instruments in real
time. During surgery, they monitor critical parameters, detect compli-
cations, and alert the team if intervention is needed. AI agents assist
decision-making, provide dynamic feedback, and automate tasks like
suturing or tissue manipulation under supervision. With continuous
learning from operations, these agents rene their models, enhancing
robotic performance and advancing safer, more efcient, and person-
alized surgical care.
In a recent study
104
highlights the emergence of AI-agent surgical
assistance, introducing multi-agent AI systems that simulate surgical
roles, enhance decision-making, and optimize workow coordination.
These intelligent systems leverage LLMs and memory-augmented AI to
provide real-time guidance, anticipate procedural steps, and facilitate
seamless team collaboration. Expanding on this, SUFIA,
105
an
LLM-driven robotic assistant agent, translates natural language com-
mands into high-level surgical plans and low-level control actions. It
integrates real-time perception modules for dynamic adaptation and
ensures safety via a HITL mechanism. In simulated experiments, SUFIA
achieved a 100 % success rate in needle lifting and 90 % in needle
handover, while in physical trials, success rates were 100 % and 50 %,
respectively. These results highlight agentic AI’s potential to enhance
surgical dexterity, optimize workow, and support autonomous yet
supervised robotic interventions. Operating within an interactive envi-
ronment, these AI agents rene surgical strategies, assist with intra-
operative navigation, and adapt dynamically to evolving scenarios. By
enhancing precision, safety, and efciency, agent-driven approaches in
robotic surgery broaden the scope of automation, paving the way for
more autonomous, adaptive surgical systems in the future.
Table 2 shows the categorization of AI agent types in healthcare,
along with key applications, users, and technologies that are associated
with each type.
3. Challenges and recommendations for agentic AI in healthcare
Implementing agentic AI systems in healthcare presents several
critical challenges that must be addressed for successful integration and
safe deployment.
3.1. Model availability and data privacy
One of the most critical challenges in healthcare AI is the availability
of diverse, high-quality data for train AI models.
106
Data availability is
restricted by regulatory barriers, fragmented healthcare infrastructures,
and stringent privacy requirements. Furthermore, medical records
contain unstructured text and sensitive data, necessitating the use of
Explainable AI (XAI) to ensure compliance with strict medical
regulations.
Recent research highlights the importance of ne-tuning LLMs to
extract structured knowledge from pathology reports while maintaining
interpretability and compliance with evolving regulatory standards. In
this context, a study
107
applied the Bidirectional Encoder Representa-
tions from Transformers (BERT) model to the medical domain, enabling
the generation of decision scores and improving pathology report an-
notations. By analyzing the resulting contextual embeddings, the study
provides valuable insights into the organization of diagnostic informa-
tion, enhancing alignment with clinical workows and enabling inte-
gration with imaging data. While LLMs improve structured knowledge
extraction, privacy-enhancing technologies, such as federated
learning
108
and differential privacy,
109
are emerging solutions, but their
implementation is resource-intensive and technically complex. Addi-
tionally, ensuring data traceability and compliance with evolving reg-
ulatory frameworks, such as the European In Vitro Diagnostic
Regulation (IVDR), is essential for AI adoption in medical diagnostics.
110
Bias in AI-driven healthcare models can lead to discriminatory de-
cisions, disproportionately affecting underrepresented groups.
Addressing this requires diverse data curation, bias mitigation, and
continuous fairness evaluations in clinical settings. Addressing this
challenge requires diverse dataset curation, bias-mitigation techniques
in model training, and ongoing fairness evaluations in real-world clin-
ical deployments.
3.2. Regulatory and compliance complexity
Healthcare regulations lag behind the rapid advancements in AI
technology. In agentic AI models, continuous learning, which adapts
based on new data, poses unique challenges to regulatory approval
processes, as their performance and risk proles may evolve over time.
Table 2
Categorization of AI agent types in healthcare with key applications, users, and technologies.
AI agents Key applications Healthcare categories Main users Key AI technologies
Image base agents Disease diagnosis, early detection,
report generation
Diagnosis, Clinical decision support Radiologists, Doctors Computer vision (CNNs, ViTs), MLLMs for
image-text integration
predictive analytics
agents
Risk prediction, disease progression
forecasting, patient outcomes
Clinical Decision Support, Treatment and
Patient Care, Drug Discovery & Research
Doctors, Care Teams Predictive Modeling, including supervised
ML, ensemble methods, and time-series
analysis
Conversational
agents
Symptom checking, patient triage,
virtual consultations
Patient Engagement and Monitoring Patients, General
Practitioners
NLP, Dialogue Systems, Pretrained LLMs
NLP agents Processing clinical notes,
summarizing EHRs, extracting
insights
Operations and Administration, Clinical
Decision Support
Medical Coders,
Analysts
NLP, Pretrained LLMs
Rule base agents Following clinical guidelines,
alerting for drug interactions
Clinical Decision Support Doctors, Pharmacists Rule-Based Reasoning, leveraging logic
programming, expert rules, knowledge graphs
Hybrid agents Combining imaging, text, video,
and predictive analytics for
decisions
Clinical Decision Support, Diagnosis,
Robot-Assisted Surgery
Doctors,
Radiologists,
Surgeons
Multimodal Learning
ML agents Disease classication, anomaly
detection, treatment planning
Diagnosis, Treatment and Patient Care,
Drug Discovery & Research
Data Scientists,
Doctors
ML/DL algorithms, RL
Expert system
agents
Emulating clinical expertise for
diagnosis and planning
Treatment and Patient Care, Clinical
Decision Support, Robot-Assisted Surgery
Specialists,
Researchers,
Surgeons
Knowledge-based systems, rule-based systems
Recommender
agents
Suggesting diagnostic tests,
personalized treatments
Treatment and Patient Care, Clinical
Decision Support
Doctors, Care Teams Collaborative ltering, recommendation
systems, RL
N. Karunanayake
Informatics and Health 2 (2025) 73–83
79
Additionally, LLMs present risks of hallucinating information,
111
potentially leading to clinical errors. Under IVDR, AI-driven diagnostics
must demonstrate scientic validity, analytical robustness, and clinical
reliability to meet compliance standards.
110
Establishing robust valida-
tion and monitoring frameworks for these evolving systems is critical to
ensuring patient safety. As a best practice, continuous model validation
and performance monitoring should be conducted. Furthermore, sur-
veillance protocols must be introduced specically for these AI models
to track decision-making accuracy.
3.3. Integration with healthcare workows
Seamlessly incorporating agentic AI into existing clinical workows
is a formidable challenge. Many healthcare institutions rely on legacy
EHRs and operational systems that are not designed to accommodate
novel AI tools. Successful integration requires addressing these technical
constraints through change management, user training, and iterative
collaboration between clinicians, IT teams, and AI developers.
Furthermore, as highlighted in AI-supported medical imaging, explain-
ability and causability are key factors in ensuring trust and adoption
among healthcare professionals, allowing them to understand and
validate AI-driven recommendations.
110
A lack of trust and acceptance among clinicians and patients remains
a major barrier to AI adoption in healthcare. To address this, AI agentic
models should undergo rigorous real-world validation, and healthcare
professionals should receive structured training on AI-assisted decision-
making. Explainability mechanisms, such as interpretable AI models and
condence scores, can also improve clinician trust and enable informed
decision-making.
3.4. Resource and infrastructure limitations
Deploying agentic AI systems at scale requires signicant computa-
tional resources and infrastructure. Many resource-constrained settings,
such as rural clinics, lack the hardware and connectivity needed to run
advanced AI systems. Moreover, the energy demands of LLMs raise
sustainability concerns.
112
Optimizing AI systems for low-resource en-
vironments is essential for equitable access to their benets.
High implementation costs remain a major challenge, especially for
small hospitals and clinics with limited budgets. To promote equitable
AI adoption, scalable and cost-effective AI solutions—such as open-
source medical AI models, cloud-based AI platforms, and public-
private funding initiatives can help reduce the nancial burden while
ensuring accessibility.
3.5. Adversarial weaknesses
AI systems in healthcare are vulnerable to adversarial attacks,
113
where maliciously crafted inputs can manipulate their outputs. This
poses signicant risks in AI agents, particularly in high-stakes scenarios
such as diagnostics, treatment planning, and robot-assisted surgeries.
Developing robust defenses and conducting rigorous testing against
these threats are critical for ensuring reliability. Strengthening AI de-
fenses will not only protect patient outcomes but also foster trust in the
adoption of agentic AI across healthcare applications.
3.6. Ethical and legal responsibilities
The autonomy of agentic AI and complexity of the data
65
in
healthcare raises signicant ethical and legal concerns, particularly
regarding liability for adverse outcomes. Assigning responsibility for
AI-related medical errors is complex due to the opaque nature of many
AI models. Establishing liability frameworks and standardized
accountability measures is essential for ethical AI deployment.
114
The
lack of transparency in how AI systems generate medical recommen-
dations makes it difcult to assign accountability in cases of
misdiagnosis or inappropriate treatment suggestions.
115
To address these challenges, legal frameworks and governance
mechanisms must be established to dene clear accountability struc-
tures. Regulatory bodies, such as the EU AI Act
116
and FDA
117
emphasis
rigorous documentation, monitoring, and human oversight to ensure
accountability. Furthermore, the agentic AI models outputs should have
a explainability mechanism, such as condence score and interpretable
outputs, that helps the clinicians understand and validate the outcome.
3.7. Human oversight and AI governance
Another major challenge in deploying agentic AI in healthcare is the
diminishing feasibility of human oversight, particularly as AI models
become more complex and autonomous. The opacity of these models
and the scale at which AI systems operate make real-time human
monitoring increasingly difcult. Recent studies highlight the need for
hybrid oversight mechanisms, incorporating HITL frameworks, rule-
based interventions, and XAI to maintain regulatory compliance and
trust in AI-driven decisions.
118
However, challenges such as automation
bias, oversight fatigue, and regulatory gaps complicate governance ef-
forts, necessitating interdisciplinary solutions that balance autonomy
with liability.
While agentic AI presents signicant challenges in healthcare,
addressing these issues through robust governance, transparency, and
clinician collaboration will enable safer and more effective AI-driven
healthcare solutions. By integrating ethical oversight, regulatory
adaptability, and technical innovations, AI can transform patient care
while maintaining high standards of safety and accountability.
4. Future directions for agentic AI in healthcare
Agentic AI in healthcare is still in its early stages, requiring rigorous
development to ensure its safety, reliability, and seamless integration
into healthcare management and clinical practice. The next phase of
agentic AI will be driven by advancements in self-evolving AI architec-
tures, multimodal integration, and real-time adaptability, enabling these
systems to assist clinicians rather than replace them. As AI continues to
evolve at an unprecedented pace, its adoption in healthcare is expected
to expand signicantly in the coming years.
94
A key research priority will be the development of Generalist AI
Agents capable of performing complex, multidisciplinary tasks with
minimal human intervention. Few-shot and self-supervised learning
techniques will play a crucial role in enabling these models to rapidly
adapt to new medical scenarios while reducing dependence on large,
annotated datasets. However, the central challenge remains: how can
these technological advancements be effectively integrated into daily
clinical workows?
To address this, hybrid human-AI collaboration frameworks will
continue to evolve, fostering deeper physician-AI interaction while
ensuring trust, interpretability, and accountability. Explainable AI and
ethical AI governance will remain focal areas, with an increasing
emphasis on auditability, transparency, and regulatory oversight to
mitigate risks in clinical decision-making. Moreover, the rapid ad-
vancements in open-source LLMs
119
and reasoning mechanisms will
drive the next milestone in the evolution of agentic AI, enhancing its
reasoning capabilities and contextual understanding in healthcare
applications.
A key direction for future work in agentic AI is transforming its
reactive nature into a proactive one. This shift will enable AI agents to
engage more effectively with clinicians and patients, bridge knowledge
gaps, and suggest optimal workows.
From an implementation standpoint, edge AI solutions will facilitate
real-time AI inference on medical devices, reducing reliance on
centralized cloud-based infrastructure and improving AI accessibility in
resource-limited environments. Additionally, federated learning and
blockchain are poised to revolutionize secure, decentralized medical
N. Karunanayake
Informatics and Health 2 (2025) 73–83
80
data sharing, ensuring interoperability while preserving patient privacy.
In the long term, agentic AI will likely evolve into universal AI
healthcare ecosystems, where AI seamlessly integrates with EHRs, im-
aging platforms, and robotic systems to enable proactive, personalized,
and preventive medicine. However, realizing this vision will require
sustained interdisciplinary collaboration among AI researchers, clini-
cians, and policymakers to ensure that agentic AI remains ethical,
transparent, clinically reliable, and aligned with patient-centered care.
5. Conclusion
The emergence of agentic AI, powered by advancements in MLLMs,
represents a transformative milestone in the healthcare industry. This
technology has the potential to revolutionize modern healthcare by
addressing longstanding challenges associated with complex and diverse
medical data. The autonomy and human-like reasoning capabilities of
agentic AI systems can mitigate inefciencies and augment decision-
making, paving the way for more precise diagnostics, personalized
treatment plans, and enhanced healthcare administration.
To fully realize its potential, collaboration between healthcare pro-
fessionals and computer scientists is essential. Such interdisciplinary
efforts can ensure that agentic AI solutions are aligned with clinical
workows and optimized for real-world application. These systems can
support a wide range of use cases, from diagnosis and treatment plan-
ning to robotic-assisted surgeries and administrative automation.
Agentic AI can be deployed in various congurations, including private
hospital-owned systems, public inter-hospital networks, or clinician-
specic tools, offering exibility and adaptability to different health-
care settings.
Importantly, this technology holds promise for bridging healthcare
disparities, particularly in low-resource environments. By enhancing
clinician efciency and diagnostic accuracy, agentic AI can signicantly
improve patient outcomes even in underprivileged areas. However, the
current developmental stage of agentic AI also brings challenges,
particularly its "blackbox" nature and unpredictable behaviors. Rigorous
research is necessary to address these concerns, including robust vali-
dation, interpretability, and transparency in AI decision-making.
Additionally, the integration of agentic AI into healthcare must be
guided by strategic legal frameworks and ethical standards to ensure
patient safety and societal trust. As we enter the "agentic era," the
adoption of these next-generation AI systems offers a promising future
for digital healthcare, with the potential to redene patient care and
operational efciency at a global scale.
Ethical Approval
This article does not contain any studies with human participants or
animals performed by the author.
Funding
The author declares that no additional funding was received to
support the work reported in this paper.
CRediT authorship contribution statement
nalan karunanayake: Writing – review & editing, Writing – original
draft.
Declaration of Generative AI and AI-assisted technologies in the
writing process
During the preparation of this work the author used ChatGPT to
assist with language editing. After using this tool, the author reviewed
and edited the content as needed and take full responsibility for the
content of the publication.
Declaration of Competing Interest
The author declares that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Acknowledgments
This research was funded in part through the NIH/NCI Cancer Center
Support Grant P30 CA008748. The content is solely the responsibility of
the authors and does not necessarily represent the funding sources.
Consent to Participate
Not applicable.
Consent to Publish
Not applicable.
References
1. Hirani R, et al. Articial intelligence and healthcare: a journey through history,
present innovations, and future possibilities. Art. no. 5 Life. May 2024;14(5).
https://doi.org/10.3390/life14050557.
2. Jiang F, et al. Articial intelligence in healthcare: past, present and future. Stroke
Vasc Neurol. Dec. 2017;2(4). https://doi.org/10.1136/svn-2017-000101.
3. van Melle W. MYCIN: a knowledge-based consultation program for infectious
disease diagnosis. Int J Man-Mach Stud. May 1978;10(3):313–322. https://doi.org/
10.1016/S0020-7373(78)80049-2.
4. Miller RA, McNeil MA, Challinor SM, et al. The INTERNIST-1/QUICK medical
reference project—status report. West J Med. Dec. 1986;145(6):816–822.
5. Barnett GO, Cimino JJ, Hupp JA, Hoffer EP. DXplain: an evolving diagnostic
decision-support system. JAMA. Jul. 1987;258(1):67–74. https://doi.org/
10.1001/jama.1987.03400010071030.
6. Kaul V, Enslin S, Gross SA. History of articial intelligence in medicine. Gastrointest
Endosc. Oct. 2020;92(4):807–812. https://doi.org/10.1016/j.gie.2020.06.040.
7. Patel VL, et al. The coming of age of articial intelligence in medicine. Artif Intell
Med. May 2009;46(1):5–17. https://doi.org/10.1016/j.artmed.2008.07.017.
8. Yin J, Ngiam KY, Teo HH. Role of articial intelligence applications in real-life
clinical practice: systematic review. J Med Internet Res. Apr. 2021;23(4), e25759.
https://doi.org/10.2196/25759.
9. Tang Y-X, et al. Automated abnormality classication of chest radiographs using
deep convolutional neural networks. Npj Digit Med. May 2020;3(1):1–8. https://
doi.org/10.1038/s41746-020-0273-z.
10. Truhn D, Schrading S, Haarburger C, et al. Radiomic versus convolutional neural
networks analysis for classication of contrast-enhancing lesions at
multiparametric breast MRI. Radiology. Feb. 2019;290(2):290–297. https://doi.
org/10.1148/radiol.2018181352.
11. Bernal J, et al. Deep convolutional neural networks for brain image analysis on
magnetic resonance imaging: a review (C) Artif Intell Med. Apr. 2019;95:64–81.
https://doi.org/10.1016/j.artmed.2018.08.008.
12. Yamashita R, Nishio M, Do RKG, Togashi K. Convolutional neural networks: an
overview and application in radiology. Art. no. 4 Insights Imaging. Aug. 2018;9(4).
https://doi.org/10.1007/s13244-018-0639-9.
13. Miotto R, Wang F, Wang S, et al. Deep learning for healthcare: review,
opportunities and challenges. Brief Bioinf. Nov. 2018;19(6):1236–1246. https://
doi.org/10.1093/bib/bbx044.
14. Bertl M, et al. Challenges for AI in Healthcare Systems. In: Steffen B, ed. Bridging the
Gap Between AI and Reality. Cham: Springer Nature Switzerland; 2025:165–186.
https://doi.org/10.1007/978-3-031-73741-1_11.
15. Holzinger A, Saranti A, Angerschmid A, et al. Toward human-level concept
learning: pattern benchmarking for AI algorithms. Patterns. Jul. 2023;4(8),
100788. https://doi.org/10.1016/j.patter.2023.100788.
16. Peters U. Explainable AI lacks regulative reasons: why AI and human decision-
making are not equally opaque. AI Ethics. Aug. 2023;3(3):963–974. https://doi.
org/10.1007/s43681-022-00217-w.
17. Mehandru N, Miao BY, Almaraz ER, et al. Evaluating large language models as
agents in the clinic. Npj Digit Med. Apr. 2024;7(1):1–3. https://doi.org/10.1038/
s41746-024-01083-y.
18. Russell SJ, Norvig P. Articial intelligence: a modern approach. Third edition, Global
edition. Prentice Hall series in articial intelligence. Boston Columbus
Indianapolis: Pearson; 2016.
19. Petrova-Dimitrova VS. Classications of intelligence agents and their applications.
Art. no. 1, 2022, Accessed: Jan. 15 J Fundam Sci Appl. 2025;28(1). 〈https://journal
s.tu-plovdiv.bg/index.php/journal/article/view/559〉. Art. no. 1, 2022, Accessed:
Jan. 15.
20. H. Touvron et al., LLaMA: Open and Efcient Foundation Language Models. Feb.
27, 2023, arXiv: arXiv:2302.13971. doi: 10.48550/arXiv.2302.13971.
N. Karunanayake
Informatics and Health 2 (2025) 73–83
81
21. E. Almazrouei et al., The Falcon Series of Open Language Models. Nov. 29, 2023,
arXiv: arXiv:2311.16867. doi: 10.48550/arXiv.2311.16867.
22. L. Zheng et al., Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. Dec.
24, 2023, arXiv: arXiv:2306.05685. doi: 10.48550/arXiv.2306.05685.
23. J. Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language
Models. Jan. 10, 2023, arXiv: arXiv:2201.11903. doi: 10.48550/arXiv.2201.11903.
24. “[2210.03629] ReAct: Synergizing Reasoning and Acting in Language Models.”
Accessed: Mar. 11, 2025. [Online]. Available: 〈https://arxiv.org/abs/221
0.03629〉.
25. S. Yao et al., Tree of Thoughts: Deliberate Problem Solving with Large Language
Models. Dec. 03, 2023, arXiv: arXiv:2305.10601. doi: 10.48550/
arXiv.2305.10601.
26. Kiryati N, Landau Y. Dataset growth in medical image analysis research. Art. no. 8
J Imaging. Aug. 2021;7(8). https://doi.org/10.3390/jimaging7080155.
27. Matengo W, Otsieno E, Wanjiru K. Big data analytics in healthcare. In: Chaari L, ed.
Digital Health in Focus of Predictive, Preventive and Personalised Medicine. Cham:
Springer International Publishing; 2020:129–138. https://doi.org/10.1007/978-3-
030-49815-3_15.
28. Refsum C, Britnell M. The missing millions: leadership and the global workforce crisis in
healthcare. Research Handbook on Leadership in Healthcare. Edward Elgar
Publishing; 2023:75–99. Accessed: Jan. 12, 2025. [Online]. Available:
〈https://www.elgaronline.com/edcollchap/book/9781800886254/book-part-978
1800886254-12.xml〉.
29. Unige, Healthcare Human Resource Shortfall in Vietnam Compared to Select
Countries in Asean. Archives Des Sciences. Accessed: Jan. 12, 2025. [Online].
Available: 〈https://unige.org/volume-74-issue-3-2024/healthcare-human-resour
ce-shortfall-in-vietnam-compared-to-select-countries-in-asean-2/〉.
30. Ng KH, Faust O, Sudarshan V, Chattopadhyay S. Data overloading in medical
imaging: emerging issues, challenges and opportunities in efcient data
management. J Med Imaging Health Inf. Aug. 2015;5(4):755–764. https://doi.org/
10.1166/jmihi.2015.1449.
31. Keni S. Evaluating articial intelligence for medical imaging: a primer for
clinicians. Br J Hosp Med. Jul. 2024;85(7):1–13. https://doi.org/10.12968/
hmed.2024.0312.
32. Nijor S, Gokcen E, Lad N. Patient safety issues from information overload in
electronic medical records: a systematic review. Res Sq. Apr. 21, 2020. https://doi.
org/10.21203/rs.3.rs-23424/v1.
33. Nijor S, Rallis G, Lad N, Gokcen E. Patient safety issues from information overload
in electronic medical records. J Patient Saf. Sep. 2022;18(6), e999. https://doi.org/
10.1097/PTS.0000000000001002.
34. Raghavendra U, et al. Brain tumor detection and screening using articial
intelligence techniques: current trends and future perspectives. Comput Biol Med.
Sep. 2023;163, 107063. https://doi.org/10.1016/j.compbiomed.2023.107063.
35. Khan MdSI, et al. Accurate brain tumor detection using deep convolutional neural
network. Comput Struct Biotechnol J. Jan. 2022;20:4733–4745. https://doi.org/
10.1016/j.csbj.2022.08.039.
36. Weikert T, Akinci D’Antonoli T, Bremerich J, et al. Evaluation of an AI-powered
lung nodule algorithm for detection and 3D segmentation of primary lung tumors,
2019 Contrast Media Mol Imaging. 2019;1:1545747. https://doi.org/10.1155/
2019/1545747.
37. Karunanayake N, Lohitvisate W, Makhanov SS. Articial life for segmentation of
fusion ultrasound images of breast abnormalities. Pattern Recognit. Nov. 2022;131,
108838. https://doi.org/10.1016/j.patcog.2022.108838.
38. Karunanayake N, et al. Dual-stage AI model for enhanced CT imaging: precision
segmentation of kidney and tumors. Art. no. 1 Tomography. Jan. 2025;11(1).
https://doi.org/10.3390/tomography11010003.
39. Kumar S, Mankame DP. Optimization driven deep convolution neural network for
brain tumor classication. Biocybern Biomed Eng. Jul. 2020;40(3):1190–1204.
https://doi.org/10.1016/j.bbe.2020.05.009.
40. Kondylakis H, et al. Documenting the de-identication process of clinical and
imaging data for AI for health imaging projects. Insights Imaging. May 2024;15(1):
130. https://doi.org/10.1186/s13244-024-01711-x.
41. Hiredesai AN, Martinez CJ, Anderson ML, et al. Is articial intelligence the future
of radiology? Accuracy of ChatGPT in radiologic diagnosis of upper extremity bony
pathology, 15589447241298982 Hand N Y N. Dec. 2024. https://doi.org/
10.1177/15589447241298982.
42. Huppertz MS, et al. Revolution or risk?-Assessing the potential and challenges of
GPT-4V in radiologic image interpretation. Eur Radio. Oct. 2024. https://doi.org/
10.1007/s00330-024-11115-6.
43. C´
amara J, Troya J, Burgue˜
no L, Vallecillo A. On the assessment of generative AI in
modeling tasks: an experience report with ChatGPT and UML. Softw Syst Model.
Jun. 2023;22(3):781–793. https://doi.org/10.1007/s10270-023-01105-5.
44. Tanno R, et al. Collaboration between clinicians and vision–language models in
radiology report generation. Nat Med. Nov. 2024:1–10. https://doi.org/10.1038/
s41591-024-03302-1.
45. Wang S, et al. Articial intelligence in lung cancer pathology image analysis. Art.
no. 11 Cancers. Nov. 2019;11(11). https://doi.org/10.3390/cancers11111673.
46. Kim SH, et al. Human-AI collaboration in large language model-assisted brain MRI
differential diagnosis: a usability study. Eur Radio. Mar. 2025. https://doi.org/
10.1007/s00330-025-11484-6.
47. M. Moor et al., Med-Flamingo: a Multimodal Medical Few-shot Learner. Jul. 27,
2023, arXiv: arXiv:2307.15189. doi: 10.48550/arXiv.2307.15189.
48. Bai F, Du Y, Huang T, et al. M3D: advancing 3D medical image analysis with multi-
modal large language models. Mar. 31. arXiv: arXiv. 2024;2404:00578. https://doi.
org/10.48550/arXiv.2404.00578.
49. Mohsen F, Al-Absi HRH, Yousri NA, et al. A scoping review of articial
intelligence-based methods for diabetes risk prediction. Npj Digit Med. Oct. 2023;6
(1):1–15. https://doi.org/10.1038/s41746-023-00933-5.
50. Huang Y-J, Chen C, Yang H-C. AI-enhanced integration of genetic and medical
imaging data for risk assessment of Type 2 diabetes. Nat Commun. May 2024;15(1):
4230. https://doi.org/10.1038/s41467-024-48618-1.
51. Deepa DrR, Sadu VB, C PG, Sivasamy DrA. Early prediction of cardiovascular
disease using machine learning: unveiling risk factors from health records. AIP
Adv. Mar. 2024;14(3), 035049. https://doi.org/10.1063/5.0191990.
52. Ayoub C, et al. Multimodal fusion articial intelligence model to predict risk for
MACE and myocarditis in cancer patients receiving immune checkpoint inhibitor
therapy. JACC Adv. Jan. 2025;4(1), 101435. https://doi.org/10.1016/j.
jacadv.2024.101435.
53. Swan M, Kido T, Roland E, dos Santos RP. AI Health Agents: Pathway2vec,
ReectE, category theory, and longevity. Art. no. 1 Proc AAAI Symp Ser. May 2024;
3(1). https://doi.org/10.1609/aaaiss.v3i1.31249.
54. Yang X. The applications of articial intelligence in personalized medicine. Appl
Comput Eng. Aug. 2024;71:47–51. https://doi.org/10.54254/2755-2721/71/
20241625.
55. Zhang H, Xi Q, Zhang F, et al. Application of deep learning in cancer prognosis
prediction model, 15330338231199287 Technol Cancer Res Treat. Jan. 2023;22.
https://doi.org/10.1177/15330338231199287.
56. Iqbal MJ, et al. Clinical applications of articial intelligence and machine learning
in cancer diagnosis: looking into the future. Cancer Cell Int. May 2021;21(1):270.
https://doi.org/10.1186/s12935-021-01981-1.
57. Hu X, et al. Interpretable medical image Visual Question Answering via multi-
modal relationship graph learning. Med Image Anal. Oct. 2024;97, 103279. https://
doi.org/10.1016/j.media.2024.103279.
58. Bai M, Yu X, Wang Y, et al. Enhancing pixel-level analysis in medical imaging
through visual instruction tuning: introducing PLAMi. Vis Comput. 2024. https://
doi.org/10.1007/s00371-024-03742-3.
59. Labkoff S, et al. Toward a responsible future: recommendations for AI-enabled
clinical decision support. J Am Med Inform Assoc. Nov. 2024;31(11):2730–2739.
https://doi.org/10.1093/jamia/ocae209.
60. Finkelstein J, Gabriel A, Schmer S, et al. Identifying facilitators and barriers to
implementation of AI-assisted clinical decision support in an electronic health
record system. J Med Syst. Sep. 2024;48(1):89. https://doi.org/10.1007/s10916-
024-02104-9.
61. A. Hoopes, V.I. Butoi, J.V. Guttag, and A.V. Dalca, VoxelPrompt: A Vision-
Language Agent for Grounded Medical Image Analysis. Oct. 10, 2024, arXiv: arXiv:
2410.08397. doi: 10.48550/arXiv.2410.08397.
62. N. Yildirim et al., Multimodal Healthcare AI: Identifying and Designing Clinically
Relevant Vision-Language Applications for Radiology. presented at the
PROCEEDINGS OF THE 2024 CHI CONFERENCE ON HUMAN FACTORS IN
COMPUTING SYTEMS, CHI 2024, 2024. doi: 10.1145/3613904.3642013.
63. Li C, et al. LLaVA-med: training a large language-and-vision assistant for biomedicine in
one day. in Proceedings of the 37th International Conference on Neural Information
Processing Systems, in NIPS ’23. Red Hook, NY, USA: Curran Associates Inc; Dec.
2023:28541–28564.
64. Oh Y, et al. LLM-driven multimodal target volume contouring in radiation
oncology. Nat Commun. Oct. 2024;15(1):9186. https://doi.org/10.1038/s41467-
024-53387-y.
65. Bertl M, Ross P, Draheim D. Systematic AI support for decision-making in the
healthcare sector: obstacles and success factors. Health Policy Technol. Sep. 2023;12
(3), 100748. https://doi.org/10.1016/j.hlpt.2023.100748.
66. Koumakis L. Deep learning models in genomics; are we there yet? Comput Struct
Biotechnol J. Jan. 2020;18:1466–1473. https://doi.org/10.1016/j.
csbj.2020.06.017.
67. Golriz Khatami S, Mubeen S, Bharadhwaj VS, et al. Using predictive machine
learning models for drug response simulation by calibrating patient-specic
pathway signatures. Npj Syst Biol Appl. Oct. 2021;7(1):1–9. https://doi.org/
10.1038/s41540-021-00199-1.
68. Kang J, Schwartz R, Flickinger J, Beriwal S. Machine learning approaches for
predicting radiation therapy outcomes: a clinician’s perspective. Int J Radiat Oncol.
Dec. 2015;93(5):1127–1135. https://doi.org/10.1016/j.ijrobp.2015.07.2286.
69. Mason DM, et al. Optimization of therapeutic antibodies by predicting antigen
specicity from antibody sequence via deep learning. Nat Biomed Eng. Jun. 2021;5
(6):600–612. https://doi.org/10.1038/s41551-021-00699-9.
70. Zhang X, et al. Deep learning with radiomics for disease diagnosis and treatment:
challenges and potential. Front Oncol. Feb. 2022;12. https://doi.org/10.3389/
fonc.2022.773840.
71. Nova SN, Rahman MdS, Hosen ASMS. Deep learning in biomedical devices:
perspectives, applications, and challenges. In: Kaiser MS, Mahmud M, Al Mamun S,
eds. Rhythms in Healthcare. Singapore: Springer Nature; 2022:13–35. https://doi.
org/10.1007/978-981-19-4189-4_2.
72. Muhammad Arslan M, Yang X, Zhang Z, et al. Advancing healthcare monitoring:
integrating machine learning with innovative wearable and wireless systems for
comprehensive patient care. IEEE Sens J. Sep. 2024;24(18):29199–29210. https://
doi.org/10.1109/JSEN.2024.3434409.
73. “AgentClinic: a multimodal agent benchmark to evaluate AI in simulated clinical
environments.” Accessed: Mar. 12, 2025. [Online]. Available: 〈https://arxiv.org/
html/2405.07960v1〉.
74. Shaik T, et al. Remote patient monitoring using articial intelligence: current state,
applications, and challenges. WIREs Data Min Knowl Discov. 2023;13(2), e1485.
https://doi.org/10.1002/widm.1485.
N. Karunanayake
Informatics and Health 2 (2025) 73–83
82
75. Nadarzynski T, Miles O, Cowie A, Ridge D. Acceptability of articial intelligence
(AI)-led chatbot services in healthcare: a mixed-methods study,
2055207619871808 Digit Health. Jan. 2019;5. https://doi.org/10.1177/
2055207619871808.
76. Catherine AT, Towfek SK, A. Abdelhamid A. An overview of the evolution and
impact of chatbots in modern healthcare services. Dec. 2023 Mesop J Artif Intell
Health. 2023:71–75. https://doi.org/10.58496/MJAIH/2023/014.
77. Christopoulou SC. Machine learning models and technologies for evidence-based
telehealth and smart care: a review. Art. no. 1 BioMedInformatics. Mar. 2024;4(1).
https://doi.org/10.3390/biomedinformatics4010042.
78. Andrikopoulou E. Chapter 13 - The rise of AI in telehealth. In: Freeman AM,
Bhatt AB, eds. Emerging Practices in Telehealth. Academic Press; 2023:183–207.
https://doi.org/10.1016/B978-0-443-15980-0.00011-9.
79. J. Li et al., Agent Hospital: A Simulacrum of Hospital with Evolvable Medical
Agents. Jan. 17, 2025, arXiv: arXiv:2405.02957. doi: 10.48550/arXiv.2405.02957.
80. Fitzpatrick KK, Darcy A, Vierhile M. Delivering cognitive behavior therapy to
young adults with symptoms of depression and anxiety using a fully automated
conversational agent (Woebot): a randomized controlled trial. JMIR Ment Health.
Jun. 2017;4(2), e7785. https://doi.org/10.2196/mental.7785.
81. S. Gabriel, I. Puri, X. Xu, et al, Can AI Relate: Testing Large Language Model
Response for Mental Health Support. Oct. 07, 2024, arXiv: arXiv:2405.12021. doi:
10.48550/arXiv.2405.12021.
82. Ahmed Z, Mohamed K, Zeeshan S, Dong X. Articial intelligence with multi-
functional machine learning platform development for better healthcare and
precision medicine. Jan. 2020 Database. 2020:baaa010. https://doi.org/10.1093/
database/baaa010.
83. Alowais SA, et al. Revolutionizing healthcare: the role of articial intelligence in
clinical practice. BMC Med Educ. Sep. 2023;23(1):689. https://doi.org/10.1186/
s12909-023-04698-z.
84. Aminizadeh S, et al. Opportunities and challenges of articial intelligence and
distributed systems to improve the quality of healthcare service. Artif Intell Med.
Mar. 2024;149, 102779. https://doi.org/10.1016/j.artmed.2024.102779.
85. Patil S, Shankar H. Transforming healthcare: harnessing the power of AI in the
modern era. Art. no. 2 Int J Multidiscip Sci Arts. Apr. 2023;2(2). https://doi.org/
10.47709/ijmdsa.v2i1.2513.
86. Rathore Y, Sinha U, Haladkar JP, et al. Optimizing patient ow and resource
allocation in hospitals using AI. 2023 Int Conf Artif Intell Innov Healthc Ind
(ICAIIHI). Dec. 2023:1–6. https://doi.org/10.1109/
ICAIIHI57871.2023.10489698.
87. Bahrami S, Rubulotta F. Articial intelligence-driven translation tools in intensive
care units for enhancing communication and research. Art. no. 1 Int J Environ Res
Public Health. Jan. 2025;22(1). https://doi.org/10.3390/ijerph22010095.
88. Gonzalez-Garcia A, P´
erez-Gonz´
alez S, Benavides C, et al. Impact of articial
intelligence–based technology on nurse management: a systematic review. J Nurs
Manag. 2024;2024(1):3537964. https://doi.org/10.1155/2024/3537964.
89. Vaithiyalingam G. Bridging the gap: AI, automation, and the future of seamless
healthcare claims processing. Art. no. 2 Afr J Artif Intell Sustain Dev. Jul. 2022;2(2).
Art. no. 2.
90. Kapadiya K, et al. Blockchain and AI-empowered healthcare insurance fraud
detection: an analysis, architecture, and future prospects. IEEE Access. 2022;10:
79606–79627. https://doi.org/10.1109/ACCESS.2022.3194569.
91. Prabhod KJ. The role of articial intelligence in reducing healthcare costs and
improving operational efciency. Art. no. 2 Q J Emerg Technol Innov. Apr. 2024;9
(2). Art. no. 2.
92. Jiang LY, et al. Health system-scale language models are all-purpose prediction
engines. Nature. Jul. 2023;619(7969):357–362. https://doi.org/10.1038/s41586-
023-06160-y.
93. “GPT-4 in a Cancer Center — Institute-Wide Deployment Challenges and Lessons
Learned | NEJM AI.” Accessed: Mar. 12, 2025. [Online]. Available: 〈https://ai.
nejm.org/doi/full/10.1056/AIcs2300191〉.
94. Gangwal A, Ansari A, Ahmad I, et al. Current strategies to address data scarcity in
articial intelligence-based drug discovery: a comprehensive review. Comput Biol
Med. Sep. 2024;179, 108734. https://doi.org/10.1016/j.
compbiomed.2024.108734.
95. Gupta R, Srivastava D, Sahu M, et al. Articial intelligence to deep learning:
machine intelligence approach for drug discovery. Mol Divers. Aug. 2021;25(3):
1315–1360. https://doi.org/10.1007/s11030-021-10217-3.
96. Harrer S, Shah P, Antony B, Hu J. Articial intelligence for clinical trial design.
Trends Pharmacol Sci. Aug. 2019;40(8):577–591. https://doi.org/10.1016/j.
tips.2019.05.005.
97. Fountzilas E, Tsimberidou AM, Vo HH, Kurzrock R. Clinical trial design in the era
of precision medicine. Genome Med. Aug. 2022;14(1):101. https://doi.org/
10.1186/s13073-022-01102-1.
98. Xu J, et al. Translating cancer genomics into precision medicine with articial
intelligence: applications, challenges and future perspectives. Hum Genet. Feb.
2019;138(2):109–124. https://doi.org/10.1007/s00439-019-01970-5.
99. Dias R, Torkamani A. Articial intelligence in clinical and genomic diagnostics.
Genome Med. Nov. 2019;11(1):70. https://doi.org/10.1186/s13073-019-0689-8.
100. S. Liu et al., DrugAgent: Automating AI-aided Drug Discovery Programming
through LLM Multi-Agent Collaboration. Nov. 24, 2024, arXiv: arXiv:2411.15692.
doi: 10.48550/arXiv.2411.15692.
101. J. Gottweis et al., Towards an AI co-scientist. Feb. 26, 2025, arXiv: arXiv:
2502.18864. doi: 10.48550/arXiv.2502.18864.
102. Iftikhar M, Saqib M, Zareen M, Mumtaz H. Articial intelligence: revolutionizing
robotic surgery: review. Ann Med Surg. Sep. 2024;86(9):5401. https://doi.org/
10.1097/MS9.0000000000002426.
103. Guni A, Varma P, Zhang J, et al. Articial intelligence in surgery: the future is now.
Eur Surg Res. Jan. 2024;65(1):22–39. https://doi.org/10.1159/000536393.
104. Wu J, Liang X, Bai X, Chen Z. SurgBox: agent-driven operating room sandbox with
surgery copilot. presented at the 2024 IEEE international conference on big data
(BigData). IEEE Comput Soc. Dec. 2024:2041–2048. https://doi.org/10.1109/
BigData62323.2024.10825748.
105. Moghani M, et al. SuFIA: language-guided augmented dexterity for robotic surgical
assistants. 2024 IEEE/RSJ Int Conf Intell Robots Syst (IROS). Oct. 2024:6969–6976.
https://doi.org/10.1109/IROS58592.2024.10802053.
106. Rajpurkar P, Chen E, Banerjee O, Topol EJ. AI in health and medicine. Nat Med.
Jan. 2022;28(1):31–38. https://doi.org/10.1038/s41591-021-01614-0.
107. Kraiˇ
snikovi´
c C, Harb R, Plass M, et al. Fine-tuning language model embeddings to
reveal domain knowledge: an explainable articial intelligence perspective on
medical decision making. Eng Appl Artif Intell. Jan. 2025;139, 109561. https://doi.
org/10.1016/j.engappai.2024.109561.
108. Xu J, Glicksberg BS, Su C, et al. Federated learning for healthcare informatics.
J Healthc Inform Res. Mar. 2021;5(1):1–19. https://doi.org/10.1007/s41666-020-
00082-4.
109. Ficek J, Wang W, Chen H, et al. Differential privacy in health research: a scoping
review. J Am Med Inform Assoc. Oct. 2021;28(10):2269–2276. https://doi.org/
10.1093/jamia/ocab135.
110. Müller H, Holzinger A, Plass M, et al. Explainability and causability for articial
intelligence-supported medical image analysis in the context of the European In
Vitro Diagnostic Regulation. N Biotechnol. Sep. 2022;70:67–72. https://doi.org/
10.1016/j.nbt.2022.05.002.
111. Z. Xu, S. Jain, and M. Kankanhalli, Hallucination is Inevitable: An Innate
Limitation of Large Language Models. Jan. 22, 2024, arXiv: arXiv:2401.11817. doi:
10.48550/arXiv.2401.11817.
112. Argerich MF, Pati˜
no-Martínez M. Measuring and improving the energy efciency
of large language models inference. IEEE Access. 2024;12:80194–80207. https://
doi.org/10.1109/ACCESS.2024.3409745.
113. Finlayson SG, Bowers JD, Ito J, et al. Adversarial attacks on medical machine
learning. Science. Mar. 2019;363(6433):1287–1289. https://doi.org/10.1126/
science.aaw4399.
114. Smith H. Clinical AI: opacity, accountability, responsibility and liability. AI Soc.
Jun. 2021;36(2):535–545. https://doi.org/10.1007/s00146-020-01019-6.
115. Raposo VL. The fty shades of black: about black box AI and explainability in
healthcare. Med Law Rev. Feb. 2025;33(1):fwaf005. https://doi.org/10.1093/
medlaw/fwaf005.
116. Edwards L. The EU AI Act: a summary of its signicance and scope. Artif Intell EU AI
Act. 2021;1:25.
117. “FDA Perspective on the Regulation of Articial Intelligence in Health Care and
Biomedicine | Digital Health | JAMA | JAMA Network.” Accessed: Mar. 10, 2025.
[Online]. Available: 〈https://jamanetwork.com/journals/jama/fullarticle/
2825146〉.
118. Holzinger A, Zatloukal K, Müller H. Is human oversight to AI systems still possible?
N Biotechnol. Mar. 2025;85:59–62. https://doi.org/10.1016/j.nbt.2024.12.003.
119. 2025. “Open-Source Large Language Models in Radiology: A Review and Tutorial
for Practical Research and Clinical Deployment. Radiology.” Accessed: Feb. 03,
2025. [Online]. Available: https://pubs.rsna.org/doi/10.1148/radiol.241073.
N. Karunanayake
Informatics and Health 2 (2025) 73–83
83
