Source-linked AI summary
Large language models in healthcare and medical domain: A review
Zabir Al Nazi, Wei Peng
TL;DR
Healthcare LLM research faces the challenge of translating rapidly advancing language-model capabilities into safe, effective medical applications. This review synthesizes model development, clinical and multimodal use cases, evaluation benchmarks, comparative performance, and unresolved ethical and privacy concerns, finding broad potential alongside the need for rigorous evaluation and further research.
Problem
Healthcare LLMs must be assessed for safety and quality because AI systems can show implementation deficiencies, while access to effective models may be limited.
Method
The review surveys healthcare LLM development, applications including multimodal systems and drug discovery, evaluation benchmarks, comparative models, and open challenges.
Results
The review finds that LLMs support diverse medical applications and that comparative analyses identify strong performance across healthcare datasets and models.
Takeaways & Limitations
LLMs have transformative potential in healthcare, but realizing it requires rigorous evaluation, robust training data, bias mitigation, privacy protection, and interdisciplinary cooperation.
Takeaways & Limitations
Healthcare LLM deployment remains constrained by explainability, security and privacy, bias, hallucinations, legal frameworks, and risks involving patient data.
Abstract
from arXiv · showhide
The deployment of large language models (LLMs) within the healthcare sector has sparked both enthusiasm and apprehension. These models exhibit the remarkable capability to provide proficient responses to free-text queries, demonstrating a nuanced understanding of professional medical knowledge. This comprehensive survey delves into the functionalities of existing LLMs designed for healthcare applications, elucidating the trajectory of their development, starting from traditional Pretrained Language Models (PLMs) to the present state of LLMs in healthcare sector. First, we explore the potential of LLMs to amplify the efficiency and effectiveness of diverse healthcare applications, particularly focusing on clinical language understanding tasks. These tasks encompass a wide spectrum, ranging from named entity recognition and relation extraction to natural language inference, multi-modal medical applications, document classification, and question-answering. Additionally, we conduct an extensive comparison of the most recent state-of-the-art LLMs in the healthcare domain, while also assessing the utilization of various open-source LLMs and highlighting their significance in healthcare applications. Furthermore, we present the essential performance metrics employed to evaluate LLMs in the biomedical domain, shedding light on their effectiveness and limitations. Finally, we summarize the prominent challenges and constraints faced by large language models in the healthcare sector, offering a holistic perspective on their potential benefits and shortcomings. This review provides a comprehensive exploration of the current landscape of LLMs in healthcare, addressing their role in transforming medical applications and the areas that warrant further research and development.
1 Introduction
LLMs are presented as promising tools for processing complex medical information and supporting clinical decision-making, diagnosis, research, and patient care. The review surveys their development, healthcare applications, evaluation, and ethical challenges.
- Motivation: Clinical language-understanding advances may support decision-making, faster diagnosis, and improved patient care by helping professionals interpret records and develop tailored treatment plans.The passage frames these capabilities as potential benefits rather than established outcomes.
- Motivation: LLMs can process and synthesize large volumes of medical literature, patient records, and clinical research to help address healthcare information overload.They can automate medical-text analysis, extract insights, and apply that knowledge to healthcare and medical research.
- Model development: Healthcare-specific adaptations such as BioBERT and ClinicalBERT were introduced to address complex terminology, lexical ambiguity, and variable usage in clinical language.The healthcare setting also requires protection of patient data and attention to bias and unintended harm.
- Review scope: The review examines LLM technology, healthcare applications, model selection, performance metrics, fairness, bias mitigation, privacy, transparency, and ethics.It is intended as a guide for medical researchers and healthcare professionals selecting models for clinical requirements.
- Review structure: The review organizes its discussion around LLM foundations, healthcare applications and evaluation, and challenges including explainability, security, bias, and ethics.These topics are assigned to Sections 2, 3, and 4, respectively.
2 Review of Large Language Models
The review traces LLM development from recurrent and Transformer-based language models to foundational and multimodal systems. It explains the architectures, training characteristics, healthcare adaptations, and multimodal components that broaden their capabilities.
- Large Language Models: LLMs typically combine large parameter counts, pre-training on vast text corpora, and task-specific fine-tuning to perform diverse NLP tasks.Their scale enables them to capture intricate linguistic patterns and nuances.
- Transformers: Transformers use self-attention to model contextual relationships and long-range dependencies more effectively than earlier recurrent neural networks.Attention assigns different importance to input words during prediction, while multiple heads process different aspects in parallel.
- Healthcare language models: Healthcare-specific BERT adaptations include BioBERT and ClinicalBERT, while GPT-3, GPT-3.5, GPT-4, and Bard represent later large language model developments.GPT-3 is described as having 175 billion parameters.
- Large Foundational Models: Foundational models are trained on extensive, diverse, largely unlabeled data, generally with self-supervised learning, enabling broad task versatility.Their applications include image classification, natural language processing, and question-answering, although enterprise integration can present challenges.
- Multi-modal Language Models: Multimodal LLMs integrate text, images, audio, or video for tasks such as visual question answering and image-text interaction.A typical architecture uses an encoder, connector, LLM, and optionally a generator; connectors may be projection-, query-, or fusion-based.
- Multi-modal Language Models: Mixture-of-Experts multimodal models selectively activate specialized submodels according to the input and task, allowing dynamic adaptation to multimodal data.Experts may be specialized for image recognition or language processing.
3 Large Language Models in Healthcare and Medical Domain
This section surveys healthcare LLMs, tracing their development from foundational models to specialized systems and examining applications, explainability, future directions, benchmarks, and constraints. It highlights diverse clinical uses and substantial variation in performance across healthcare tasks.
- Development of Healthcare LLMs: Healthcare language modeling progressed from Transformer-based systems such as BERT and RoBERTa toward specialized models for medical language and applications.BioBERT and ClinicalBERT address complex terminology, lexical ambiguity, and variable usage in clinical text.
- Healthcare Applications: Healthcare LLMs support clinical decision assistance, medical record analysis, patient engagement, health information dissemination, and automated diagnosis-related tasks.The review presents these applications as potential ways to improve diagnostic accuracy, administrative efficiency, personalization, and comprehensiveness of healthcare delivery.
- Healthcare Applications: LLMs are also applied to drug discovery by analyzing molecular structures, identifying promising compounds, and forecasting efficacy and safety profiles.Warm-started biochemical language models outperformed baseline models, while the one-stage strategy generalized better in docking evaluation and benchmark metrics.
- Explainability: Explainability methods for healthcare LLMs include prompting strategies, multi-agent clinical reasoning, medical knowledge graphs, and visual-textual explanation techniques.These approaches target interpretability and reliability in mental-health analysis, clinical decision reasoning, automated diagnosis, and pathology visual question answering.
- Future Trajectories: Future directions include multimodal biomolecular understanding, real-time adaptive learning for clinical decision support, and federated learning across institutions.Federated systems are described as enabling privacy-preserving knowledge propagation without direct data sharing, while biomolecular modeling could support genomics and personalized medicine.
- Challenges and Constraints: Evaluation requires frameworks that assess safety and quality because effective AI systems can fail after implementation and some leading models lack public accessibility.Limited transparency can hinder examination of model structure and outputs, including efforts to identify and address bias and hallucinations.
- Performance Evaluation and Benchmarks: Performance varies substantially across healthcare tasks: GPT-4 (2024) reaches 93.06% accuracy on MedQA, BioELECTRA-Base reaches 86.34% on MedNLI, and elEmBERT-V1 reaches 0.961 AUC on Tox21.PubMedQA performance is led by Meditron-70B (CoT + SC) at 81.6% accuracy, followed by BioGPT-Large (1.5B) at 81.0%.
4 Limitations and Open Challenges
LLMs in healthcare face unresolved challenges involving transparency, privacy, bias, hallucinations, security, and legal accountability. Addressing these constraints requires rigorous evaluation, safeguards for sensitive data, bias mitigation, and regulatory compliance.
- Healthcare LLM integration requires explainability, security and privacy protections, bias mitigation, hallucination reduction, and clear legal frameworks.
- Opaque recommendations and biased training data can hinder professional acceptance and produce incorrect diagnoses or treatment recommendations.
- Medical research with LLMs risks patient re-identification, privacy-invading inferences, and inadvertent inclusion of personally identifiable information in pre-training data.
- LLM outputs may reproduce demographic, disease, or treatment biases, requiring rigorous validation, verification, and mitigation across data and generated outputs.
- Plausible but factually inaccurate hallucinations create risks for medical guidance and clinical decision-making, while opacity complicates content verification.
- Med-HALT evaluates medical hallucinations through reasoning-based and memory-based assessment categories.
- Ethical risks include harmful diagnoses without emotional support and indistinguishable generated text that can facilitate misinformation, plagiarism, or impersonation.
- EU AI Act, HIPAA, GDPR, and MDR promote transparency, data protection, patient safety, and accountability but increase development costs and timelines.
5 Conclusion
The review concludes that LLMs offer broad potential for clinical language understanding, decision support, and information retrieval across healthcare applications. Their wider impact depends on rigorous evaluation, robust data, bias mitigation, privacy protection, ethical standards, and continued interdisciplinary research.
- LLMs show potential for enhancing clinical language understanding and medical applications across healthcare.
- Their applications span named entity recognition through question-answering, supporting decision support and information retrieval.
- Comparisons of state-of-the-art and open-source LLMs emphasize their significance for innovation and collaboration in healthcare.
- Performance metrics support continuous improvement but require rigorous evaluation that accounts for bias and ethical concerns.
- Realizing LLMs’ potential requires robust training data, bias mitigation, data privacy, further research, interdisciplinary cooperation, and ethical application.