Large language models in healthcare and medical domain: A review

• Clinical language systems could support decision-making, expedite diagnosis, improve patient care, and help professionals interpret records and formulate tailored treatment plans.These systems are framed as aids for healthcare professionals facing an expanding body of medical knowledge.Source: §1 · p. 1
• LLMs can synthesize medical literature, patient records, and clinical research to ease information overload and extract insights that may advance research and patient care.Healthcare data is complex, heterogeneous, and large in scale.Source: §1 · p. 1
• GPT-3.5, GPT-4, and Bard have demonstrated proficiency across natural-language understanding tasks relevant to healthcare communication and information processing.Source: §1 · p. 2
• BioBERT and ClinicalBERT [14] [15] adapt BERT to clinical language, while healthcare use of LLMs raises concerns about ethics, privacy, security, bias, and unintended harm.The adaptations address complex terminology, lexical ambiguity, and variable usage in medical text.Source: §1 · p. 2
• The review aims to guide clinical LLM selection and examine underlying technology, healthcare applications, evaluation metrics, and challenges including fairness, bias, privacy, transparency, and ethics.Its sections cover model foundations, healthcare applications and clinical performance metrics, and challenges such as explainability, security, bias, and ethical considerations.Source: §1 · p. 2

2 Review of Large Language Models

• LLMs combine hundreds of millions to billions of parameters with broad-corpus pretraining and task-specific fine-tuning, enabling adaptation to varied NLP tasks.Pretraining on diverse datasets without task-specific annotations provides broad linguistic knowledge; smaller task-appropriate datasets are then used for fine-tuning.Source: §2 · pp. 2–3

2.1 Transformers

• Transformer self-attention weighs contextual token relationships, while parallel attention heads and positional encoding help model long-range dependencies.Multi-head attention computes attention sets in parallel, then concatenates and linearly transforms them.Source: §2.1 · p. 3, §2 · p. 3

2.2 Large Foundational Models

• Foundation models train on extensive, diverse, unlabeled data, often with self-supervision, for broad tasks, but enterprise adaptation and integration can be challenging.The task range described includes language comprehension, text and image generation, dialogue, image classification, and question-answering.Source: §2.2 · pp. 3–4

2.3 Multi-modal Language Models

• MLLMs integrate modalities such as text, images, audio, and video; Visual ChatGPT [31] and BLIP-2 [32] illustrate cross-modal interaction for tasks including visual question answering.BLIP-2 uses a Qformer to integrate visual features with textual data; iterative training can freeze some components while fine-tuning others.Source: §2.3 · p. 4
• A typical MLLM uses an encoder, connector, and language model; Mixture of Experts selectively activates specialized models to handle multimodal inputs while managing training and inference costs.Connectors may use projection-based or query-based token-level fusion, or fusion-based feature-level integration; a generator can also produce non-text outputs.Source: §2.3 · p. 5, Figure 2 · p. 5

3.1 Large Language Models for Medical and Healthcare Applications

• Healthcare language models progressed from rule-based methods and feature extraction to Transformers and biomedical models culminating in larger LLMs by 2023.The Transformer architecture led to pretrained models such as BERT and RoBERTa; the review also describes biomedical model trajectories including BioBERT, GPT-2, and MedPaLM.Source: §3 · p. 5, §3.1 · p. 6

3.2 Use Cases of Large Language Models in Healthcare

• LLMs are discussed for diagnosis, personalized patient care, clinical decision support, literature analysis, and drug discovery, including warm-started molecule generation.The reviewed molecule-generation study found warm-started models outperformed baselines, the one-stage strategy generalized better in docking evaluation and benchmark metrics, and beam search outperformed sampling for assessing compound quality.Source: §3.2 · pp. 7–8

3.3 Explainable AI Methods for Interpreting Healthcare LLMs

• Healthcare explainability work uses feature attribution, perturbation, attention, knowledge graphs, and explanation-generating frameworks, while interpretability lacks standardized evaluation metrics.Examples include SHAP, LIME [89], t-SNE, attention mechanisms [15], and knowledge graphs [91]; TraP-VQA [91] uses Grad-CAM and SHAP for visual and textual explanations.Source: §3.3 · p. 9

3.4 Future Trajectories of Large Language Models in Healthcare

• Proposed future directions include biomolecular data and real-time adaptive support, privacy-preserving federated learning, and wearable-linked multimodal health monitoring.These avenues are described for applications including genomics, personalized medicine, surgical or emergency decision support, cross-institution knowledge sharing without direct data exchange, and continuous monitoring outside clinical settings.Source: §3.4 · p. 10

3.5 Performance Evaluation and Benchmarks

• Healthcare LLM evaluation requires safety and quality frameworks and careful metric selection, while limited access to effective models hinders transparency and scrutiny of bias and hallucinations.The review lists F1 score, BLEU, GLUE, and ROGUE among standard evaluative criteria and describes the metrics as quantitative and qualitative measures of efficacy and suitability across healthcare applications.Source: §3.5 · pp. 10–11

3.6 Quantitative Performance Comparison of LLMs in Healthcare Domain

• 93.06% accuracy on MedQA makes GPT-4 (2024) the reported leader over Med-PaLM 2 (CoT + SC) (2023) at 83.7% and Meerkat-7B (Ensemble) (2024) at 74.3%.On MedNLI, BioELECTRA-Base (2021) achieves 86.34% accuracy; on Tox21, elEmBERT-V1 (2023) achieves 0.961 AUC; on PubMedQA, Meditron-70B (CoT + SC) (2023) and BioGPT-Large (1.5B) (2023) achieve 81.6% and 81.0% accuracy, respectively.Source: Figure 4 · p. 11, §3.6 · p. 12

4.1 Model Explainability and Transparency

• Black-box opacity obscures how recommendations are produced, undermining clinician acceptance, trust, and validation; clinicians should review outputs while interpretable frameworks remain a challenge [129] [130].The cited work calls for transparency and interpretable frameworks in medical imaging and clinical risk prediction, and cautions professionals to validate LLM recommendations before clinical use.Source: §4.1 · p. 12

4.2 Security and Privacy Considerations

• Privacy risks include PII in pretraining, sensitive-attribute inference, and re-identification from even anonymized health data; anonymization, secure storage, monitoring, and governance are needed [131].The passages also call for technical protections against re-identification and patient and provider involvement in development to support transparency and trust.Source: §4.2 · pp. 12, 14

4.3 Bias and Fairness

• LLMs can perpetuate or amplify gender, racial, and other training-data biases, affecting diagnoses, treatment recommendations, and equitable healthcare outcomes.The passages recommend rigorous output validation, careful data curation, routine audits, and collaboration among domain experts, ethicists, and data scientists to identify and mitigate bias.Source: §4.3 · p. 14

4.4 Hallucinations and Fabricated Information

• Plausible but inaccurate hallucinations can cause serious harm in medical guidance or clinical decisions, and increasingly persuasive outputs without source links are difficult to verify [133].Med-HALT evaluates medical hallucinations using reasoning-based and memory-based tests of problem-solving and information retrieval [43].Source: §4.4 · p. 15

4.5 Legal and Ethical Reasons

• Ethical risks include distressing diagnoses without emotional support and blurred authorship that enables misinformation, plagiarism, and impersonation; regulation can increase deployment costs and timelines [135].The cited discussion names the EU AI Act, HIPAA, GDPR, and MDR as frameworks affecting healthcare AI deployment, with requirements including documentation, encryption, access controls, and clinical testing.Source: §4.5 · p. 15

5 Conclusion

• LLMs offer versatile healthcare applications, from named entity recognition to question-answering, supporting decision support and information retrieval.The conclusion describes their potential to enhance clinical language understanding and medical applications.Source: §5 · pp. 15–16
• Comparisons of state-of-the-art and open-source LLMs highlight their significance, while performance metrics support improvement and call for rigorous evaluation standards.Source: §5 · p. 16
• Robust training data, bias mitigation, data privacy, and ethical standards remain challenges, requiring further research and interdisciplinary cooperation.The conclusion calls for collective efforts to ensure responsible application and improved patient care.Source: §5 · p. 16