Truace
Subscribe

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 betwee…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2209· cite & verify
AI-Powered Structuring of Clinical Information in Parkinson's Disease: An Approach Based on Large Language Models
Pyogenic infective diseases of the brain and spinal cord [electronic resource] : meningitis abscess of brain, infective sinus thrombosis by MacEwen, William, Sir University College, London. Library Services. Public domain

In brief

Researchers built a pipeline in which GPT-4o reads de-identified free-text neurology records from a tertiary service and fills a structured schema of 22 binary indicators of advanced Parkinson's disease, plus the diagnosis date. The records covered 1,201 confirmed patients identified between 2018 and 2023.

The work matters because many features that define advanced Parkinson's disease sit in narrative notes, which limits large-scale phenotyping and longitudinal analysis. The study reports that structured output can support research and modeling, but precision was modest, diagnosis dates were weak, and the authors call for targeted clinical review of critical fields.

Main points

  1. The retrospective cohort comprised 1,201 confirmed Parkinson's patients from a tertiary neurology service, identified between January 2018 and January 2023.
  2. The schema aligned with domains derived from the CDEPA questionnaire and used 22 binary indicators plus the diagnosis date.
  3. Repeated sampling with vote-based aggregation, recursive routing, and deterministic temporal normalization were used to reduce variability.
  4. Four neurologists annotated 120 cases as the reference standard, covering 2,760 variable instances.
  5. Precision was 61.4%, well below the 94.6% negative predictive value.

The gain

A schema-guided LLM pipeline with vote-based aggregation extracted 22 advanced Parkinson's indicators from 1,201 patients' records, with 85.4% sensitivity and 94.6% NPV against neurologist annotation.

The rundown

The pipeline has four functional blocks: presence, absence, or unknown classification; controlled-option selection; temporal-period extraction; and date extraction. Records were de-identified with a privacy-preserving step before GPT-4o processed them.

Against the expert reference, the method outperformed zero-shot biomedical and clinical baselines and direct GPT-4o. Paired patient-level analyses showed significant improvements over GPT-4o and the no-voting ablation on all six metrics after Holm correction.

The authors say diagnosis-date extraction remained the main limitation. They propose using the workflow with targeted clinical review for critical fields rather than as a fully autonomous tool.

Sources

  1. Peer-reviewedIEEE Journal of Biomedical and Health Informatics2026-10-21

The debate