Real-World Usage Patterns of Large Language Models in Healthcare

ObjectiveTo characterize real-world LLM use by healthcare professionals and identify gaps between actual usage and research focus. Materials and MethodsWe analyzed chat interactions from a secure deployment of GPT-3.5/4 at an academic medical center (Dec 2023-May 2024), classifying tasks using GPT-4o-mini according to a published taxonomy. ResultsAmong 25,173 interactions from 3,913 users, 64.1% were healthcare-related. Most common tasks were writing for professional communication (23.9%), enhancing medical knowledge (12.5%), general writing support (9.7%), and medical research (8.7%). Note-taking (1.3%) and billing/coding (0.35%) were rare. Task frequencies correlated moderately with literature (r = 0.75). DiscussionReal-world usage diverged from literature in key areas. Writing support and medical research tasks were prevalent in practice yet underexplored in research. Clinical decision support, note-taking, and billing showed limited adoption despite being promising applications, suggesting workflow and implementation barriers. ConclusionThese findings help hospital administrators align LLM deployment with actual needs and address implementation barriers.