Large language models for patient-centred evidence generation and use

Author preprint. Not peer reviewed. For patients with several health conditions, evidence for one disease may leave important questions about care unanswered. Large language models (LLMs) offer ways to help formulate these questions, examine the literature, and prepare requests for research. We propose an evidence operating system (Evidence OS) that follows a patient's question from clinical goals through evidence synthesis and research to the return of findings. The framework applies across specialties, with multimorbidity illustrating the difficulties of judging what evidence means for an individual. LLM assistance would require source checks and professional review. Similarity between patients and fluent explanations cannot establish treatment benefit. Prospective evaluation should examine supported answers, clinical use, workload, patient experience, and health outcomes. Comparisons with and without LLM assistance are needed to distinguish model effects from better organisation of evidence services. Whether the proposed framework improves care remains to be tested.