Large Language Models for Clinical Data Extraction in Workplace Injury Rehabilitation: Protocol for a Retrospective Pilot Study of Accuracy, Fairness, and Methodological Considerations in Workers' Compensation Medical Chart Review.

چکیده اصلی

BACKGROUND: Electronic medical charts within Workplace Safety and Insurance Board (WSIB) specialty programs contain rich but unstructured clinical and sociodemographic data essential for injury classification, treatment planning, and compensation decisions. Manual chart review is time-consuming, inconsistent, and susceptible to reviewer bias. Large language models (LLMs) extract complex information from unstructured clinical text, yet their application to workplace injury rehabilitation remains unexplored. OBJECTIVE: This study will evaluate the accuracy, fairness, and methodological implications of using LLMs to extract clinical and rehabilitative information from WSIB medical charts at Trillium Health Partners (THP), Canada. We will assess model performance, examine algorithmic bias across demographic subgroups, and explore ethical implications for clinical decision-making and workplace compensation. METHODS: We will conduct a retrospective review of 50 medical charts from the WSIB Back and Neck specialty program at THP, spanning January 2018 to December 2024. General-purpose models (Qwen3-VL-8B and InternVL3.5-8B) and domain-specific biomedical models (MedGemma-27B and LLaMA-3-Meditron-8B) will be evaluated off the shelf using zero-shot and few-shot prompting; the biomedical models will also be fine-tuned. Charts will be partitioned at the chart level into development, training, and held-out test sets, preventing leakage across purposes. Ground truth will be established by independent human annotation of all 50 charts, with interannotator agreement quantified using Cohen κ. The primary outcome is the macroaveraged F1-score across categorical variables on the held-out test set; secondary outcomes are the per variable F1-score, mean absolute error and root mean squared error for continuous variables, and span-level F1-score. Progression to the larger study requires a macroaveraged F1-score of at least 0.80, a pragmatic feasibility criterion; fairness and continuous-variable results are supporting outcomes. Algorithmic fairness will be examined using demographic parity and equalized odds across subgroups. A secure hybrid architecture will keep all identifiable personal health information within the THP infrastructure, with cloud compute restricted to transient processing. RESULTS: This study was funded by the Data Sciences Institute at the University of Toronto in April 2025; the award supported protocol development and ended on April 30, 2026. As of August 12, 2026, neither had a research ethics application been submitted nor had any chart been accessed. Applications to the THP and University of Toronto research ethics boards are anticipated in late 2026, and no data will be retrieved or processed before approval from both boards. Contingent on approval and further funding, data collection is anticipated through 2027, analysis in late 2027 to early 2028, and results submitted for publication in 2028. CONCLUSIONS: This protocol describes the first systematic evaluation of LLM-based data extraction applied to workplace injury medical charts. Findings will inform best practices for the responsible, reproducible, and equitable integration of these tools into occupational health and workers' compensation decision-making. INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID): PRR1-10.2196/99807.

متن کامل اصلی