Early Prediction of Parkinson's Disease Progression by Integrating Research Cohort and Real-World Data Using Knowledge-Anchored Graph Learning

Parkinson disease (PD) progression is highly heterogeneous. Deeply phenotyped longitudinal research cohorts have enabled characterization of PD progression trajectories. Early prediction of these progression patterns can help us better understand patient disease conditions and manage appropriately. However, the sample sizes of these cohorts are typically too small to build robust early predictors, and it is usually challenging to translate them to real-world patients because of differences in the population as well as information availability. In this paper we present MedStitcher, a graph-based machine learning framework that stitches individual multimodal data across research cohorts and real-world data (RWD) using a biomedical knowledge graph-anchored architecture. This design enables predictive modeling under modality missingness and cross-dataset population heterogeneity. On the research cohort data combining PPMI and PDBP, MedStitcher achieved an AUROC of 0.819 +/- 0.040 for predicting rapid PD progressors, outperforming existing machine learning approaches. Graph-based model interpretation revealed clinical and molecular signals involving cognitive vulnerability, alpha-synuclein biology, vesicle trafficking and neuroinflammation. Importantly, MedStitcher-predicted rapid progressors in the RWD cohort demonstrated elevated risks of dementia, falls, mild cognitive impairment, and gait impairment, which also enabled identification of early indicators of rapid PD progression in real-world patient populations.