Performance, Failures, and Oversight of a Large Language Model Agent for Clinical Data Analysis: Evaluation Study

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Large language model (LLM) agents capable of generating and executing statistical code from natural language may broaden access to clinical data analysis, yet which pipeline stages they perform reliably and which require expert oversight remain poorly defined. Objective This study aimed to evaluate the performance and systematic failure modes of an LLM agent across 5 stages of a clinical data analysis workflow. Methods The publicly available dataset and R script (R Foundation for Statistical Computing) were drawn from a previously published study of 12-year outcomes in 7802 patients with eyes with neovascular age-related macular degeneration at Moorfields Eye Hospital. Participants were evaluated using an LLM agent (Claude; Anthropic) across 3 interaction modes (Chat, Code, and Cowork). It was asked to perform 3 levels of data analysis practice: prompt A, to generate research questions from raw data only; prompt B, to develop a statistical analysis plan (SAP) from a high-level clinical objective, then execute it; and prompt C, to execute an analysis given an investigator-drafted SAP. Each was replicated 3 times (27 total runs). Qualitative evaluation of research question thematic coverage (prompt A), SAP completeness against a reference checklist (prompt B), and evaluation of execution outputs against validated reference values and of result text and narrative summaries against execution logs (prompts B and C) was conducted. Results The agent generated 18 clinically grounded questions spanning 7 domains; Cowork mode uniquely reached 3 thematic areas requiring data-driven methods. All 9 SAPs correctly identified the statistical framework. Kaplan-Meier estimates were near-identical across 17 completed runs. Systematic execution errors emerged: SAP quality did not predict code correctness, and within-mode errors propagated identically across independent repetitions. Result text accurately reflected execution logs in nearly all runs, though unit propagation and an undisclosed postcrash rerun were identified. Of 17 narrative summaries, 8 were fully satisfactory; 2 runs produced clinically meaningful errors. Conclusions LLM agents perform reliably for question generation and SAP drafting but require expert verification of formula composition, cohort boundary logic, and concordance computation before results are reported. Using an ophthalmology dataset as a controlled testbed, this study develops and applies an evaluation framework whose lessons are likely applicable across clinical specialties.