PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey of Large Language Models in Medicine: Progress, Application, and Challenge

Hongjian Zhou, Fenglin Liu, Boyang Gu, Xinyu Zou, Jinfa Huang, Jinge Wu, Yiru Li, Sam S. Chen, Peilin Zhou, Junling Liu, Yining Hua, Chengfeng Mao, Chenyu You, Xian Wu, Yefeng Zheng, Lei Clifton, Zheng Li, Jiebo Luo, David A. Clifton

arXiv:2311.05112v7cs.CLcs.AI

TL;DR

Existing medical LLM research has produced promising results but lacks practical guidance for development and often overlooks clinical utility. This review synthesizes how medical LLMs are built, evaluated, deployed, and improved, reporting strong benchmark performance while identifying challenges in real-world use.

  • Problem

    Existing reviews emphasize clinical applications, especially commercial LLMs, without providing practical guidelines for developing medical LLMs.

  • Method

    The review organizes evidence on medical LLM development, evaluation, clinical deployment, challenges, and optimization, including comparisons with traditional models and deployment guidelines for seven clinical scenarios.

  • Results

    86.5 and 90.2 accuracy were achieved by medical LLMs compared with 87.013 for human experts on the USMLE.

  • Takeaways & Limitations

    The review provides a practical resource for selecting development strategies, including parameter-efficient fine-tuning and prompting-based approaches, for medical applications.

  • Takeaways & Limitations

    Medical datasets remain small relative to general-purpose LLM datasets, limiting coverage of medical knowledge and contributing to weaker performance on real-life tasks.

Abstract

from arXiv · show

Large language models (LLMs), such as ChatGPT, have received substantial attention due to their capabilities for understanding and generating human language. While there has been a burgeoning trend in research focusing on the employment of LLMs in supporting different medical tasks (e.g., enhancing clinical diagnostics and providing medical education), a review of these efforts, particularly their development, practical applications, and outcomes in medicine, remains scarce. Therefore, this review aims to provide a detailed overview of the development and deployment of LLMs in medicine, including the challenges and opportunities they face. In terms of development, we provide a detailed introduction to the principles of existing medical LLMs, including their basic model structures, number of parameters, and sources and scales of data used for model development. It serves as a guide for practitioners in developing medical LLMs tailored to their specific needs. In terms of deployment, we offer a comparison of the performance of different LLMs across various medical tasks, and further compare them with state-of-the-art lightweight models, aiming to provide an understanding of the advantages and limitations of LLMs in medicine. Overall, in this review, we address the following questions: 1) What are the practices for developing medical LLMs 2) How to measure the medical task performance of LLMs in a medical setting? 3) How have medical LLMs been employed in real-world practice? 4) What challenges arise from the use of medical LLMs? and 5) How to more effectively develop and deploy medical LLMs? By answering these questions, this review aims to provide insights into the opportunities for LLMs in medicine and serve as a practical resource. We also maintain a regularly updated list of practical guides on medical LLMs at https://github.com/AI-in-Health/MedLLMsPracticalGuide

1 Introduction

This review examines how medical LLMs are developed, evaluated, deployed, and improved, addressing gaps in practical guidance and clinical applicability. It combines model-development analysis, cross-task evaluation, clinical-use guidelines, challenge assessment, and future research directions.

  • Motivation and scope: Medical LLMs emerged by adapting general LLMs such as PaLM, LLaMA, GPT, and ChatGLM to medicine.Examples include MedPaLM-2, ChatDoctor, MedAlpaca, PMC-LLaMA, BenTsao, and Clinical Camel.
  • Motivation and scope: Existing reviews emphasize clinical applications but provide limited practical guidance for developing medical LLMs.Applications discussed include EHRs, discharge summaries, health education, and care planning.
  • Review contributions: The review analyzes medical LLM architectures, parameter scales, and development datasets to support models tailored to computational, privacy, and knowledge-base requirements.It organizes development around pre-training, fine-tuning, and prompting, while also summarizing model architectures.
  • Review contributions: It evaluates medical LLMs across ten biomedical NLP tasks and compares their capabilities with traditional AI approaches.The tasks include both discriminative and generative tasks.
  • Review contributions: The review develops guidelines for seven clinical application scenarios and examines hallucination, ethical, legal, safety, and regulatory challenges.It frames trustworthy evaluation as necessary for responsible clinical use.
  • Future directions: Future directions emphasize interdisciplinary co-development of training data, benchmarks, metrics, and deployment strategies to support responsible clinical integration.The proposed collaboration connects AI specialists with medical professionals.

2 The Principles of Medical Large Language Models

Medical LLMs are developed through pre-training, fine-tuning, or prompting, using varied architectures and medical data sources. These approaches trade off domain adaptation, computational cost, and task performance.

  • Development approaches: Medical LLMs are mainly developed through pre-training, fine-tuning, or prompting general models for the medical domain.Table 2 organizes models according to development method and summarizes their parameters, data scale, and sources.
  • Pre-training: Pre-training uses structured and unstructured medical corpora, including EHRs, clinical notes, medical literature, PubMed, MIMIC-III, and PMC.Models may combine multiple corpora or use specialized sources such as clinical practice guidelines.
  • Pre-training: Pre-training refines objectives such as masked language modeling, next sentence prediction, and next token prediction according to model architecture.BERT-series models mainly use masked language modeling and next sentence prediction, whereas GPT-series models mainly use next token prediction.
  • Fine-tuning: Fine-tuning adapts general LLMs with supervised, instruction, or parameter-efficient methods using medical conversations, question-answer pairs, and knowledge graphs.SFT adds medical-domain training, IFT uses instruction-input-output triples, and PEFT includes LoRA, Prefix Tuning, and Adapter Tuning.
  • Fine-tuning: Parameter-efficient fine-tuning keeps most original model parameters frozen while adding small trainable components, reducing computational demands while maintaining performance.LoRA adds trainable low-rank matrices, Prefix Tuning adds continuous task-specific vectors, and Adapter Tuning inserts small neural modules.
  • Prompting: Prompting can achieve strong medical-task results without training model parameters, while retrieval-augmented generation grounds answers in retrieved medical documents.MedPrompt combines prompting strategies, and RAG systems retrieve relevant guideline or document content using embedding-based similarity.

3 Medical Tasks

Medical LLMs support discriminative and generative medical tasks, but their strongest performance is concentrated in close-ended question answering rather than open-ended clinical tasks.

  • Discriminative Tasks: Medical tasks include discriminative activities such as question answering, entity extraction, relation extraction, classification, inference, similarity, and retrieval.Their inputs include clinical questions, notes, documents, research papers, and patient EHRs; outputs may be labels, entities, relationships, or answers.
  • Generative Tasks: Generative tasks produce new medical text for summarization, text generation, and simplification.Applications include condensing radiology findings, generating discharge instructions or health advice, and clarifying complex medical language.
  • Performance Comparisons: GPT-4 consistently outperforms task-specific fine-tuned models on evaluated medical QA datasets and is comparable to human experts.The evaluated QA datasets include MedQA (USMLE), PubMedQA, and MedMCQA.
  • Performance Comparisons: BioBERT achieves an F1 score of 89.36 versus GPT-4’s 56.73 on NCBI disease entity extraction.The comparison illustrates that general LLMs can underperform task-specific lightweight models on non-QA discriminative tasks.
  • Performance Comparisons: Current general LLMs are comparable to experts on exam-style close-ended QA but remain weaker on open-ended non-QA tasks without answer options.The review therefore calls for evaluation across a broader range of medical tasks before clinical decision-making integration.

4 Clinical Applications

Medical LLMs are being explored across clinical applications, but most remain in research and development, with limited real-world validation and no large-scale clinical trials specifically targeting these models.

  • Clinical Applications: Current medical LLM research explores clinical decision support, report generation, medical robotics, and multilingual medical information.These efforts target recommendations, report drafting, decision-making and diagnostic support, surgical collaboration, and translation for foreign patients.
  • Clinical Applications: Table 3 organizes medical LLMs by architecture, development method, parameter count, pre-training and fine-tuning data, data sources, tasks, and performance.The table is intended to guide model selection, construction, and evaluation for different clinical applications.
  • Clinical Applications: Researchers are evaluating medical LLM effectiveness and safety across healthcare settings, but large-scale clinical trials specifically targeting these models are missing.Clinical expansion is described as subject to rigorous validation processes and ethical reviews.

4.1 Medical Decision-Making

LLMs offer opportunities for medical decision-making by processing information from patient histories, clinical data, and medical literature, while specialized knowledge integration can support diagnosis and treatment suggestions.

  • Medical Decision-Making: Medical decision-making includes diagnosis, prognosis, treatment suggestion, risk prediction, and clinical trial matching.These activities require synthesizing and interpreting information from patient histories, clinical data, and current medical literature.
  • Medical Decision-Making: Effective medical decision-making systems should combine a robust LLM with specialized medical knowledge.The review describes integrating UMLS knowledge graphs and fine-tuning T5 models as an example.
  • Medical Decision-Making: LLMs used alone for diagnosis rely heavily on subjective patient text and cannot directly analyze medical diagnostic imagery.Image-based systems such as ChatCAD instead use an external computer-aided diagnosis model to convert image outputs into text for LLM-based reasoning.

4.2 Clinical Coding

LLMs are being applied to clinical coding, especially ICD coding, to automate extraction and code assignment, but hallucinated or unsupported codes and limited coverage beyond ICD remain important challenges.

  • Clinical Coding: LLMs can extract medical terms from clinical notes and assign ICD, medication, and procedure codes.The intended benefits include reducing manual workload and improving coding accuracy and consistency.
  • Clinical Coding: PLM-ICD fine-tunes RoBERTa for ICD coding and reports strong performance on 70,539 MIMIC-II and MIMIC-III notes.Other approaches include DRG-LLaMA, ChatICD, and LLM-codex; the latter trains an LSTM on ChatGPT responses.
  • Clinical Coding: ICD coding is usually formulated as multi-label classification and evaluated with F1, AUC, and Precision@k, often using MIMIC-III.Evaluations may consider either the top k most frequent labels or the full label set.
  • Clinical Coding: Generative LLMs may hallucinate codes outside the candidate list or invent nonexistent clinical codes when processing lengthy inputs.The review emphasizes proactive error detection and correction before incorrect codes enter patient EHRs.
  • Clinical Coding: Most clinical-coding research focuses on ICD coding because of its widespread use and large datasets such as MIMIC-III.The review identifies medication and procedure coding as areas needing broader LLM coverage.

4.3 Clinical Report Generation

LLMs support clinical report generation through text summarization and multimodal image-text methods, with evaluation relying on both general and radiology-specific metrics. Despite potential efficiency gains, hallucinations, verbosity, and limitations of lexical evaluation remain important concerns.

  • LLMs can assist clinicians by summarizing clinical documentation, improving efficiency and reducing potential errors in lengthy reports.
  • Vision-language systems combine medical-image annotations with textual prompts to generate reports that follow specified parameters and structures.
  • Radiology report generation includes unimodal summarization and multimodal image-text summarization, with approaches such as dynamic prompting, iterative optimization, and LoRA adaptation.
  • MIMIC-III and MIMIC-IV are commonly used for evaluation, alongside lexical metrics, BERTScore, and radiology-specific measures such as CheXbert similarity, RadGraph, and RadCliQ.
  • LLM-generated reports may be more comprehensive and precise than human reports, but can hallucinate, interpret inputs literally, and remain less concise.
  • Lexical metrics can inaccurately assess clinical report quality because semantically equivalent sentences may use different wording.

4.4 Medical Robotics

LLMs are being integrated into medical robotics to augment decision-making, communication, interaction, and control. However, clinical evaluation and safe operation remain constrained by simulated testing, procedural complexity, and risks from errors and biases.

  • LLMs complement medical robotics by augmenting decision-making, communication, interaction, and control capabilities.
  • Representative systems illustrate LLM integration into robotic surgery and other medical-robotics applications, despite challenges from healthcare complexity and real-world evaluation.
  • Clinical evaluations commonly rely on simulated scenarios, controlled laboratory settings, and expert human assessment rather than comprehensive testing in actual healthcare environments.
  • Errors and biases in route planning and motion control can cause robots to misjudge human intentions or make inappropriate decisions in shared workspaces.
  • Sensing technologies and physical design constraints are proposed to reduce the occurrence and consequences of judgment errors in human-robot environments.

4.5 Medical Language Translation

Medical language translation covers terminology translation and dialogue simplification, with multilingual fine-tuning and retrieval augmentation supporting communication across professional and lay audiences. Misinterpretation and discriminatory wording remain consequential deployment risks.

  • Medical language translation addresses terminology conversion and dialogue interpretation to support communication among professional and non-professional groups.
  • Fine-tuning pre-trained models on parallel medical corpora, including scientific articles, clinical notes, and glossaries, supports multilingual terminology translation.
  • Dialogue simplification benefits from training on technical conversations paired with lay-language explanations, with retrieval augmentation helping provide accessible wording.
  • Evaluation uses accuracy for some multiple-choice benchmarks and BLEU, ROUGE, METEOR, and BERTScore for generative benchmarks, supplemented by domain-specific criteria.
  • Misinterpretation in translation and simplification can have damaging consequences, while discriminatory wording may cause miscommunication and legal consequences.

4.6 Medical Education

LLMs are used in medical education for explanation, translation, question answering, exam preparation, tutoring, and public-facing simplification. Evaluation should combine quantitative and qualitative methods, while bias and hallucination risks require careful verification.

  • LLMs support medical education through explanations, translation, question answering, exam preparation, and Socratic-style tutoring.
  • LLMs can tailor medical explanations to different audiences by varying the degree of technical terminology used in generated text.
  • Existing pre-trained models such as ChatGPT and Med-Gemini can provide knowledge synthesis, question answering, content generation, and explanations for self-study.
  • Medical-education evaluation can combine assessment or clinical-skills outcomes with surveys and focus groups involving students and educators.
  • Biases in training data and insufficient ethical training can propagate stereotypes, discrimination, and outputs inconsistent with professional ethical principles.
  • Hallucinations can produce plausible but incorrect information that misleads learners and may contribute to misconceptions, inappropriate treatment strategies, or misdiagnosis.

4.7 Mental Health Support

LLMs are being explored for mental health support because conversational interaction may improve access and disclosure, but emotional understanding and harmful responses remain critical concerns.

  • Motivation: LLMs may lower access barriers to mental health support through conversational companionship and reduced treatment costs.Many psychotherapies rely on lengthy patient-doctor conversations that can be cost-prohibitive.
  • Potential benefits: Patient willingness to discuss mental health topics with robots may support diagnosis and treatment, alongside convenience and lower financial stakes.
  • Guideline: Mental health systems can begin with existing LLMs and specialize them using medical question-answering or counseling dialogue data.Examples include PsyChat, ChatCounselor, and Psy-LLM, with evaluations combining automated and human assessment.
  • Discussion: LLMs may struggle to understand complex emotional states and provide the empathy and human connection required in therapeutic interactions.
  • Discussion: Improperly trained or controlled LLMs may produce insensitive or harmful advice that conflicts with evidence-based psychological practices and mental health guidelines.Mitigation requires training in evidence-based practices, ethical considerations, and risk assessment protocols.

4.8 Medical Inquiry and Response

Medical inquiry systems use LLMs to assist diagnostic conversations and personalized recommendations, but reliability, workflow integration, privacy, and validation remain barriers to real-world deployment.

  • Motivation: Medical LLMs are being investigated to answer patient inquiries and assist physicians with documentation as demand for accessible healthcare grows.
  • Guideline: LLM-based consultation systems combine dialogue, memory, and processing components to support healthcare professionals and patient care.These systems are intended to provide diagnostic conversations and personalized recommendations beyond rule-based algorithms or limited datasets.
  • Evaluation: Evaluation commonly uses accuracy, precision, recall, and F1-score, supplemented by assessments of inquiry capability, fluency, accuracy, and safety.Comparisons may involve human experts or established models such as ChatGPT.
  • Discussion: Real-world deployment remains distant because biased or inaccurate outputs could cause improper advice or misdiagnosis.Reliability and generalizability require rigorous testing across diverse patients and medical contexts.
  • Discussion: Deployment also requires substantial technical and organizational integration, alongside careful protection of patient privacy and data security.
  • Ethics and responsibility: Transparency, explainability, accountability, informed consent, and the doctor-patient relationship require ongoing attention in AI-assisted healthcare.Collaboration among researchers, healthcare professionals, ethicists, and policymakers is identified as necessary.

5 Challenges

Medical LLM adoption is constrained by hallucinations, inadequate evaluation, limited and inaccessible data, difficult knowledge updates, alignment gaps, and unresolved ethical and regulatory concerns.

  • 5.1 Hallucination: Hallucinations can produce inaccurate or unverifiable medical information, causing misdiagnoses, inappropriate treatments, and harmful patient education.They include intrinsic contradictions of facts and extrinsic outputs such as fabricated citations or evasive answers.
  • 5.1 Hallucination: Training-time, generation-time, and retrieval-augmented correction are the main categories of approaches proposed to mitigate hallucinations.Training-time methods adjust model weights, while other methods add reasoning or external knowledge during generation.
  • 5.2 Lack of Evaluation Benchmarks and Metrics: Medical QA benchmarks such as MedQA and MedMCQA do not adequately measure trustworthiness, helpfulness, explainability, or faithfulness.The review therefore calls for domain-specific and LLM-specific benchmarks and metrics.
  • 5.2 Lack of Evaluation Benchmarks and Metrics: TruthfulQA and HaluEval assess truthfulness but do not cover the medical domain, leaving a need for more medically relevant evaluation resources.
  • 5.3 Lack of Medical Data: Small medical datasets cover only a limited portion of medical knowledge, contributing to weaker performance on differential diagnosis and personalized treatment planning than on open benchmarks.
  • 5.3 Lack of Medical Data: Medical data access is restricted by ethical, legal, and privacy procedures, while labeling is hindered by scarce experts and small margins of error.Synthetic data may broaden coverage, but training on generated datasets has been reported to cause forgetting and requires further validation.
  • 5.4 New Knowledge Adaptation: Knowledge updates must address both removing outdated knowledge without unintended associations and delivering new information in real time.These barriers are especially consequential where accurate and timely medical knowledge is required.
  • 5.4 New Knowledge Adaptation: Model editing and retrieval-augmented generation are proposed for knowledge adaptation, but model-editing effectiveness varies across architectures.

6 Future Directions

Future medical LLM research must extend beyond early-stage general-language applications toward clinically realistic evaluation, multimodal systems, collaborative agents, specialized fields, and interdisciplinary governance.

  • 6 Future Directions: Medical LLM integration remains in its infancy, leaving many development and deployment avenues for researchers and practitioners.
  • 6.1 Evaluation: Traditional accuracy-focused medical QA benchmarks inadequately capture the broader clinical skills required for application.Passing standardized medical exams does not necessarily demonstrate proficiency in nuanced clinical tasks.
  • 6.2 Multimodal LLMs: Medical multimodal models could combine visual, audio, and language inputs, but few currently process time-series data such as ECGs and PPGs.Scaling their training remains costly.
  • 6.3 Medical Agents: LLM-based agents combine language-model reasoning with external tools and multimodal perception to interact with environments, learn from feedback, and acquire skills.
  • 6.3 Medical Agents: Collaborative medical agents could model specialist roles such as radiologists, cardiologists, and pathologists and integrate their interpretations into a holistic opinion.Medical integration remains challenging because diagnosis often requires multiple investigations and decision-making roles.
  • 6.4 Specialized Fields: General medicine dominates current medical LLM research, leaving specialized fields such as rehabilitation therapy and sports medicine under-represented.
  • 6.5 Interdisciplinary Collaboration: Collaboration between medical and technology communities is essential for scrutinizing training data, ethical protocols, privacy protection, safety, and efficacy.
Loading 2311.05112v7…