Abstract / Summary
Abstract Background: Parkinson's disease (PD) diagnosis relies primarily on clinical motor examination, and the prodromal phase preceding overt motor manifestations remains difficult to identify. No single biomarker modality has been shown to be both sensitive and specific across the full disease spectrum, and how best to integrate multimodal data for stage-specific detection is unclear. Objective: To optimize a multimodal biomarker framework using machine learning based on Parkinson's Progression Markers Initiative (PPMI) data, and to define stage-specific biomarkerpanels for manifest PD detection, prodromal identification, and longitudinal progression monitoring. Methods: We analyzed 948 PPMI participants (445 PD, 389 prodromal, 111 healthy controls [HC], 3 SWEDD) with 18 candidate biomarkers spanning dopamine transporter SPECT (DaTscanSBR), structural MRI (FreeSurfer 7 cortical and subcortical metrics), diffusion tensor imaging (DTI), and MDS-UPDRS Parts I–IV. Feature-wide association analysis with Mann-Whitney U tests and Cohen's d was followed by random forest Gini importance ranking. Random forest classifiers (500 trees, 5-fold stratified cross-validation) were trained on three feature sets: (1) DaTscan + clinical; (2) structural MRI only; (3) full multimodal combination. Four-year longitudinal trajectories (BL, V06, V10) of motor scores and grey matter volume were characterized. Results: For PD vs HC, the DaTscan-clinical model achieved near-perfect discrimination (AUC = 0.999, 95% CI 0.997–1.000), far exceeding structural MRI alone (AUC = 0.526). Top discriminative features were putaminal SBR, striatal SBR, and MDS-UPDRS Part III motor score. Prodromal vs HC discrimination was more challenging (AUC = 0.700), with subcortical grey matteratrophy as the leading signal while DaTscan SBR was non-discriminative (P = 0.43). PD vs prodromal classification was excellent (AUC = 0.984). Longitudinally, PD motor scores increased ~1.5 points/year and grey matter volume declined ~0.85%/year. Conclusions: Multimodal biomarkers display staging-dependent complementarity: DaTscan and motor scales dominate manifest PD identification, while subcortical volumetric MRI enablesprodromal detection. Stage-tailored biomarker panels support improved PD phenotyping and early screening for clinical trial enrichment.