Search

Large language models (LLMs) have rapidly advanced since the transformer architecture was introduced in 2017, with systems such as GPT-3, GPT-4, Med-PaLM, and Claude increasingly explored for applications in medical education, clinical documentation, decision support, and patient communication, raising both optimism and concerns regarding safety and reliability. This systematic review synthesizes evidence across studies retrieved from PubMed, arXiv, ACL Anthology, IEEE Xplore, and Google Scholar that empirically evaluated LLMs in clinical settings using quantitative performance metrics, with risk of bias assessed using an adapted PROBAST framework for machine learning research. Findings show that LLMs achieve 60–90% accuracy on USMLE-style examinations, with leading models such as GPT-4 and Med-PaLM 2 reaching or surpassing passing thresholds, while in clinical documentation tasks they can reduce physician workload by approximately 30–50% in generating outputs such as discharge summaries, though human review remains consistently required. Performance in clinical decision support is more variable and specialty-dependent, and hallucination rates ranging from 5–30% have been reported, alongside persistent issues of bias and overconfidence in incorrect outputs. Overall, while LLMs demonstrate strong capabilities in structured medical knowledge tasks and documentation support, current limitations including hallucinations, bias, and lack of prospective clinical validation prevent safe autonomous deployment, making clinician oversight and robust safety safeguards essential for any clinical use.

This article proposes a conceptual framework for a diagnostic support system in emergency departments that leverages large language models, retrieval-augmented generation, and chain-of-thought reasoning. By combining triage notes and vital signs, the system generates a ranked differential diagnosis list to assist clinicians without replacing their judgment. The framework includes components like a triage note encoder, a vital sign encoder, a retrieval module, and a diagnosis ranker, using evidence from clinical guidelines, curated references, and de-identified prior cases. The approach grounds the model in authoritative knowledge while ensuring transparency and explainability in the diagnostic process. However, prospective validation, integration into workflows, and clinician oversight are crucial before implementation to ensure safety and effectiveness.

Clinical trial recruitment is hindered by slow, costly, and labor-intensive processes, particularly due to the complexity of eligibility criteria often written in free text. This systematic review examines the use of large language models (LLMs) for matching clinical trial eligibility criteria to electronic health records (EHR). It evaluates zero-shot, few-shot, and fine-tuned LLM approaches, comparing their strengths, limitations, and deployment readiness in supporting patient-trial matching. Thirty-three studies published from 2017 to 2026 were included, with findings showing that zero-shot prompting is most adaptable for simple criteria, few-shot prompting offers consistent reasoning for ambiguous criteria, and fine-tuned models excel in task-specific performance but require labeled data and are less portable. The review concludes that no single approach is optimal for all trial screening tasks, and hybrid workflows combining various methods with human verification are most suitable for clinical use.