Source-linked AI summary
A Survey of LLM-based Agents in Medicine: How far are we from Baymax?
Wenxuan Wang, Zizhan Ma, Zheng Wang, Chenghan Wu, Jiaming Ji, Wenting Chen, Xiang Li, Yixuan Yuan
TL;DR
Healthcare LLM-based agents offer capabilities for clinical assistance, but their architectures, applications, evaluation, and deployment challenges require systematic synthesis. This survey reviews these dimensions across medical-agent research and concludes that reliability and clinical utility remain constrained by hallucination, multimodal, reasoning, integration, bias, privacy, and safety challenges.
Problem
Medical LLM-based agents show promise in healthcare, yet the field faces safety, implementation, ethical, multimodal, and reliability challenges requiring structured review.
Method
The survey systematically reviews medical-agent architectures, applications, evaluation frameworks, metrics, challenges, and future research directions.
Results
The survey identifies system profiles, clinical planning, medical reasoning, and external capacity enhancement as core architectural components spanning major medical applications and evaluation approaches.
Takeaways & Limitations
Future work should prioritize real-time error correction, improved multimodal fusion, and hybrid reasoning to enhance reliability and clinical utility.
Takeaways & Limitations
The review primarily covers works published between 2022 and early 2024 and focuses mainly on English-language publications in major academic databases.
Abstract
from arXiv · showhide
Large Language Models (LLMs) are transforming healthcare through the development of LLM-based agents that can understand, reason about, and assist with medical tasks. This survey provides a comprehensive review of LLM-based agents in medicine, examining their architectures, applications, and challenges. We analyze the key components of medical agent systems, including system profiles, clinical planning mechanisms, medical reasoning frameworks, and external capacity enhancement. The survey covers major application scenarios such as clinical decision support, medical documentation, training simulations, and healthcare service optimization. We discuss evaluation frameworks and metrics used to assess these agents' performance in healthcare settings. While LLM-based agents show promise in enhancing healthcare delivery, several challenges remain, including hallucination management, multimodal integration, implementation barriers, and ethical considerations. The survey concludes by highlighting future research directions, including advances in medical reasoning inspired by recent developments in LLM architectures, integration with physical systems, and improvements in training simulations. This work provides researchers and practitioners with a structured overview of the current state and future prospects of LLM-based agents in medicine.
1 Introduction
LLM-based medical agents are being applied across clinical tasks, but safety, implementation, and ethical challenges remain. This survey systematically reviews their architectures, applications, evaluation, challenges, and future directions.
- LLM-based agents support diagnostic assistance, patient communication, and medical education by combining language models with medical knowledge and healthcare systems.They are designed to understand complex medical situations and provide evidence-based recommendations.
- The field still faces implementation issues, safety concerns, and ethical considerations that affect reliable healthcare deployment.
- The survey examines system profiles, external capacity enhancement, clinical planning, and medical reasoning as core components of medical agents.
- It reviews clinical and administrative applications, evaluation frameworks, performance metrics, and challenges affecting reliability, safety, and clinical integration.
- The review analyzes 60 studies published between 2022-2024, selected from 300 initial search results and narrowed to 80 after screening.
2 Background
LLM-based medical agents extend standard LLMs with autonomous interaction, external knowledge, planning, and tool use. Deployment requires multimodal processing, accuracy, reliability, and human-AI collaboration with physician oversight.
- An LLM-based agent integrates external knowledge retrieval, task planning, and tool invocation to support structured decision-making in real-world applications.
- Unlike standard LLMs, medical agents operate autonomously and adapt dynamically to new information and tasks.
- Healthcare deployment requires agents to process and synthesize clinical text, imaging, and laboratory results for decision support.
- Agents should facilitate information sharing and human-AI collaboration while maintaining physician oversight.
3 LLM-based Medical Agent Architecture
The survey presents medical-agent architectures that organize roles, planning, reasoning, and external capabilities. These components support task decomposition, adaptive collaboration, structured inference, multimodal perception, and clinical tool use.
- Medical agents require well-defined architectures to integrate complex clinical knowledge and support safe, effective deployment.
- Agent profiles define role attributes, behavioral patterns, and operational competencies through specialized functional modules.
- Clinical Planning: Clinical planning decomposes complex medical tasks into subtasks that interact with clinical tools and databases, improving operational efficiency and error correction.
- Clinical Planning: The framework includes Single Agent, Sequential Task Chain, Collaborative Experts, and Iterative Evolution paradigms for structuring medical workflows.
- Medical Reasoning: Medical reasoning uses sequential inference, parallel hypothesis exploration, and iterative feedback to improve diagnostic precision and decision robustness.
- External Capacity Enhancement: External capacity enhancement connects agents to EHRs, OCR, medical-image models, knowledge sources, calculators, EHR interfaces, and image-analysis software.
4 Application Scenarios
LLM-based medical agents span clinical decision support, data analytics, training simulations, and healthcare service optimization. Reported applications use specialized agents, retrieval, simulation, and automation to support clinical and operational work.
- Clinical Decision Support and Diagnosis: Multi-agent frameworks assign specialized roles for intent recognition, diagnostic reasoning, treatment planning, and structured clinical discussion.Examples include MedAide, MDagents, and EHRagent, which coordinate reasoning across diagnostic and treatment stages.
- Clinical Data Analytics and Documentation: Clinical data analytics and documentation systems process structured and unstructured data using advanced architectures and retrieval-augmented generation.ColaCare performs mortality prediction and hospital-readmission analysis on MIMIC-III and MIMIC-IV.
- Medical Training and Simulation: Medical training systems simulate interactions across specialties and complex healthcare scenarios to evaluate diagnostic and treatment performance before clinical use.Agent Hospital supports repeated training through large-scale simulations.
- Healthcare Service Optimization: Healthcare service optimization automates patient education, data collection, and support services, reducing professional workload while maintaining service quality.
- Healthcare Service Optimization: These implementations have produced measurable improvements in operational efficiency and patient satisfaction.
5 Evaluation and Benchmarking
The survey organizes evaluation around benchmark types and metrics that assess medical knowledge, clinical workflows, reasoning, communication, and generated text. Static benchmarks provide exact-answer testing, while workflow simulations better represent interactive clinical decision-making.
- Evaluation frameworks are needed to assess reliability, safety, clinical effectiveness, limitations, and improvement needs across medical tasks.The survey presents evaluation metrics and benchmark categories for healthcare applications.
- Benchmark categories: Static question-answering benchmarks measure medical knowledge using predetermined answers across standardized tasks.Examples include MedQA, MedMCQA, PubMedQA, and MMLU.
- Benchmark categories: Workflow-based simulation benchmarks model multistage clinical decision-making, including multimodal analysis, dialogue, and diagnosis across specialties and diseases.Examples include MedChain, AI Hospital, AgentClinic, and ClinicalLab.
- Benchmark categories: RJUA-SPs applies automated evaluation in urology using standardized patients and retrieval-augmented techniques.
- Evaluation metrics: Exact-match metrics such as accuracy, precision, and recall compare outputs with reference answers for tasks including multiple-choice medical questions.These metrics are effective for factual knowledge but insufficient for complex reasoning or detailed explanations.
6 Discussions
The discussion identifies persistent risks and deployment barriers for medical LLM agents, including hallucinations, heterogeneous data, workflow integration, bias, privacy, and physical-system safety. It also points to reasoning improvements, richer patient participation, and more realistic training simulations as future directions.
- Hallucination Management: Hallucinations can produce incorrect or misleading medical information, creating risks of erroneous diagnoses and treatments and requiring verification and dynamic error correction.The survey highlights MedHallBench and HaluEval as benchmarks for hallucination-related risks.
- Multimodal and Multilingual Integration: Medical agents must integrate clinical text and images while handling multilingual and multicultural terminology, documentation standards, and regional practices.
- Cross-Department Integration: Different healthcare departments use distinct workflows and documentation standards, making interoperability and accurate data exchange difficult.The survey proposes universal standards and adaptive interfaces as future work.
- Evaluation Challenges: Static benchmarks miss sequential decision-making, adaptive reasoning, patient communication, multimodal interactions, clinical outcomes, and cross-setting generalizability.BLEU and ROUGE measure textual overlap rather than outcomes such as diagnostic accuracy, while dataset bias can limit generalizability.
- System Integration Complexity: Large-scale healthcare systems such as Polaris illustrate integration complexity, and broader deployment does not always improve operational efficiency.
- Implementation Barriers: Developing and maintaining medical agents requires substantial computational resources and costs, while reliability concerns can undermine cost-effectiveness.
- Patient-Centered Care: Current diagnostic systems receive limited feedback from patients and caregivers, although patient narratives and caregiver input are important for patient-centered decision-making.
7 Conclusion
The survey concludes that medical LLM agents enhance diagnostics, data analysis, and clinical workflows but remain limited by hallucinations, multimodal integration, and medical reasoning accuracy. It prioritizes real-time error correction, improved multimodal fusion, and hybrid reasoning for greater reliability and clinical utility.
- Medical LLM agents show promise across diagnostics, data analysis, and clinical workflows, while hallucination management, multimodal integration, and reasoning accuracy remain unresolved.
- Future research should pursue real-time error correction, improved multimodal fusion, and hybrid reasoning to enhance reliability and clinical utility.
Limitations
The review primarily covers publications from 2022 to early 2024 and mainly focuses on English-language studies from major academic databases. Consequently, later developments and research outside this coverage may not be represented.
- The review mainly covers works published between 2022 and early 2024, so later architectures and approaches may not be captured.
- The review focused mainly on English-language publications in major academic databases despite aiming for comprehensive coverage.
A.1 Paper Collection
The review systematically searched major academic databases for studies on LLM-based medical agents published between 2022 and 2024, narrowing about 300 identified articles to 80 shortlisted papers.
- A.1 Paper Collection: About 300 articles were initially identified through searches of PubMed, ACM Digital Library, arXiv, and Google Scholar.The search used keywords including “large language model,” “medical agent,” “clinical decision support,” and “healthcare AI.”
- A.1 Paper Collection: 80 papers were shortlisted after title and abstract screening for relevance and quality.
A.2 Evaluation and Benchmarking
The review presents Table 3 as a summary of common evaluation benchmarks and metrics for LLM-based agents in medicine.
- A.2 Evaluation and Benchmarking: Table 3 summarizes common evaluation benchmarks and metrics for LLM-based agents in medicine.