A data-driven framework for long-term risk stratification of advanced Parkinson’s disease using PPMI

Abstract

Abstract Advanced Parkinson disease has prognostic and therapeutic implications, yet staging tools are qualitative and difficult to operationalize for longitudinal modelling and cross-cohort comparison. We developed a reproducible operationalization that translates the 13-item Diagnostic Criteria for Advanced Parkinson Disease questionnaire into structured variables and generates longitudinal labels capturing certainty of advanced disease. In the Parkinson’s Progression Markers Initiative near-diagnosis cohort ( n  = 1,302; up to 13 years), we applied this pipeline to characterize label trajectories and face validity over time. As a proof of utility, we used baseline clinical and genetic features to forecast advanced disease at years 7–11, explicitly separating forecasting from contemporaneous staging. Using a binary long-horizon endpoint, the best year-9 model showed an area under the receiver operating characteristic curve of 0.89 (95% CI 0.81–0.97). In an independent real-world cohort with ≥ 11 years follow-up ( n  = 35), discrimination attenuated (0.55–0.61), consistent with dataset shift and limited event counts.