From Clinical Criteria to AI-Based Classification of Advanced Parkinson’s Disease: A Data-Driven Approach Using Structured PPMI Data

Advanced Parkinson’s disease (APD) involves severe motor and non-motor complications and requires early identification, yet lacks a standardized quantitative definition. This study translated expert-defined CDEPA criteria into 216 structured variables from the Parkinson’s Progression Markers Initiative (PPMI) dataset to train machine learning models for early APD classification. A 1,302 patients cohort was followed up for 13 years. A label-rescuing strategy addressed longitudinal incompleteness. Supervised models trained on baseline data predicted future APD status. Binary classifiers outperformed multiclass approaches; the best-performing model (XGBoost, Year 9) achieved AUC 0.881, balanced accuracy 0.824, and F1 score 0.819. Top predictors included genetic mutation status, age, MDS-UPDRS I–II, REM sleep behavior disorder, and tremor severity. Non-motor symptoms—especially autonomic dysfunction and sleep disturbances—were more informative than motor signs, comprising 57.1% of top features. These findings support the feasibility of early APD classification and propose a scalable, data-driven framework for APD prediction.