Off-the-shelf Large Language Models (LLM) Are Of Insufficient Quality To Provide Medical Treatment Recommendations, While Customization of LLMs Result In Quality Recommendations.

Abstract

The content accuracy from off-the-shelf LLMs mirror the content accuracy of the unregulated Internet from which these generative artificial intelligence models are supplied. With error rates approximating 30% in terms of treatment recommendations for the management of common musculoskeletal conditions, seeking expert opinion remains paramount. However, custom LLMs represent an excellent opportunity to infuse niche, bespoke expertise from the many specialties and subspecialties within medicine. Methods of customizing these generative models broadly fall under the categories of prompt engineering, "retrieval augmented generation" (RAG) prioritizing retrieval of relevant information from a specific domain of data, "fine-tuning" a basic pre-trained model into one that is refined for healthcare-related vernacular and acronyms, and "agentic augmentation" including software that break down complex tasks into smaller ones, recruiting multiple LLMs (with or without RAG), optimizing the output, internally deciding if the response is appropriate or sufficient, and even passing on an unmet outcome to a human for supervision ("phone a friend"). Custom LLMs offer physicians and their associated organizations the rare opportunity to regain control of our profession by reestablishing authority in our increasingly digital landscape.