AI-Powered Structuring of Clinical Information in Parkinson's Disease: An Approach Based on Large Language Models.

Advanced Parkinson's Disease (APD) is clinically heterogeneous, and many APD-defining features are recorded predominantly in unstructured neurology narratives, limiting scalable phenotyping and longitudinal analyses. An approach based on large language models (LLMs) was designed to transform free text from electronic health records (EHRs) into an APD-oriented dataset aligned with Questionnaire for the Evaluation of Advanced Parkinson's Disease (CDEPA)-derived domains. A retrospective EHR cohort identified between January 2018 and January 2023 from a tertiary neurology service comprised 1,201 confirmed Parkinson's disease (PD) patients after diagnostic eligibility filtering. Following privacy-preserving de-identification, GPT-4o populated a schema of 22 binary APD-relevant indicators and the PD diagnosis date using four functional blocks for presence/absence/unknown classification, controlled-option selection, temporal-period extraction, and date extraction. Repeated sampling with vote-based aggregation, schema-guided recursive routing, and deterministic temporal normalization were used to reduce stochastic variability and constrain temporal inference. Performance was assessed against an expert reference standard of 120 cases annotated by four neurologists. Across 2,760 variable instances, accuracy was 83.3%, precision 61.4%, sensitivity 85.4%, specificity 82.7%, negative predictive value (NPV) 94.6%, and F1-score 71.4%. The methodology achieved higher performance than all evaluated zero-shot biomedical/clinical baselines and direct GPT-4o; paired patient-level analyses also confirmed significant improvements over GPT-4o and the no-voting ablation across all six metrics after Holm correction. Diagnosis-date extraction remained the main limitation. The proposed schema-driven workflow can convert heterogeneous EHR narratives into APD-relevant structured variables for longitudinal research and downstream modeling, with targeted clinical review for critical fields.