Interpretable NLP-Based Analysis of Longitudinal Electronic Health Records for Early Prediction of Parkinson’s Disease: A Retrospective Multicenter Study in India

Parkinson’s disease (PD) is often diagnosed after substantial delay in India, limiting timely intervention. Unstructured longitudinal electronic health records (EHRs) may contain prodromal signals, but NLP-based approaches are understudied in Indian settings.We conducted a retrospective multicenter study using de-identified EHRs from AIIMS New Delhi and PGIMER Chandigarh (2015–2025). From 45,000 documents across 15,000 patients with neurological symptoms, 2,500 had confirmed PD. After exclusions and NLP pre-filtering, the final cohort comprised 8,500 patients (1,800 PD-positive, 6,700 controls). PD was defined by MDS clinical criteria; DaTSCAN supported 78% of diagnoses. Mimics were excluded. We developed an NLP-XGBoost pipeline using BioBERT for named entity recognition, negation detection, and temporal BiLSTM embeddings aggregated via attention pooling. A temporal split prevented leakage; a 6-month pre diagnosis washout excluded diagnostic workup notes. SHAP provided interpretability. The model achieved AUC 0.923 (95% CI 0.908 0.937), sensitivity 89.4%, specificity 91.2%, PPV 85.6%, NPV 93.1%, and F1-score 0.874, outperforming logistic regression (AUC 0.847) and CNN (AUC 0.891). Calibration was good (Brier 0.082; slope 1.02), PR-AUC was 0.891, and NRI was 0.32. External validation at SGPGI Lucknow yielded AUC 0.897. PD risk was predicted 2–5 years before diagnosis, with performance across horizons (AUC 0.941 at 6 months to 0.898 at 5 years). SHAP identified slow turning (mean |SHAP| 0.42), speech dysfluency (0.31), and REM sleep complaints (0.28) as top predictors. Rural patients showed bradykinesia signals 18 months earlier. Among 1,240 high-risk patients, 68% converted within 4 years (HR 12.4). Retrospective modeling projected a 41% reduction in diagnostic delay. Interpretable NLP of Indian EHRs offers a scalable, imaging-free prescreening approach with potential to support earlier referral. Prospective validation is warranted to confirm clinical impact.