Extracting TNFi switching reasons and trajectories from real-world data using large language models

To evaluate whether large language models (LLMs) can automate chart review to identify tumor necrosis factor inhibitor (TNFi) switching patterns and reasons for switching in a large real-world cohort.

We conducted an observational study using de-identified electronic health record (EHR) data from 2012 to 2023 at a single academic medical center (University of California, San Francisco). TNFi medication orders and linked clinical notes were extracted, requiring at least 6 months of follow-up to identify treatment switches, defined as a change from one TNFi to another at consecutive encounters. Using GPT-4, we extracted which TNFi was stopped and started and classified the reason for switching. Performance was benchmarked against eight open-source LLMs, structured EHR data, and expert annotation.

A total of 9187 patients (mean [SD] age, 39.9 [19.0] years; 57.1% female) received ≥1 TNFi with sufficient follow-up. We identified 3104 TNFi switches among 2112 patients. GPT-4 achieved micro-F1 scores of 0.75 (stopped drug), 0.80 (started drug), and 0.83 (reason). Among open-source models, Starling-7B-beta and Llama-3-8B performed most competitively. The most common reason identified by GPT-4 was lack of efficacy (56.9%), followed by adverse events (13.5%) and insurance/cost (10.8%).

Both GPT-4 and locally deployable LLMs effectively extracted complex treatment trajectories and rationale from clinical notes, supporting their broader utility in scalable EHR review and real-world evidence generation.