Imaging-Validated Parkinson’s Subtypes via Sensitivity Grid SuStaIn Approximation and Conditional Normalising Flow

Background: Patient stratification in Parkinson’s disease (PD) is constrained by clinical heterogeneity, longitudinal attrition, and the absence of frameworks that compare subtyping configurations under matched feature families, cluster numbers, and modelling assumptions. Existing approaches rarely validate recovered partitions with quantitative imaging biomarkers under multiple-testing control or with a generative model whose dispersion diagnostics are transparently reported. Methods: We develop a fully traceable multi-source data fusion pipeline for the Parkinson’s Progression Markers Initiative (PPMI) cohort, combining: (i) clinical subtype discovery via a z-score SuStaIn approximation explored across a 96-configuration sensitivity grid; (ii) external validation of the recovered subtypes against longitudinal DaTscan and R2* MRI biomarkers; and (iii) conditional trajectory generation via a Real NVP normalising flow. Methodology selection is performed on the n=798 longitudinally complete subset using BIC, longitudinal stability, and imaging validation as independent criteria. Results: The two-subtype Subscores partition (slow-/rapid-progression, BIC=−531,724, n=798) achieves the best information criterion, the highest longitudinal retention (80-87% across visit pairs), and the strongest imaging validation. DaTscan striatal binding ratios differentiate the two subtypes under Benjamini–Hochberg correction at V06 (q<0.001 for right caudate and bilateral putamen mean) and V10. R2* substantia nigra markers do not reach significance (n=14–51, underpowered; reported as exploratory). The conditional normalising flow converges stably (epoch 8 checkpoint, NLLval=42.41) with near-correct dispersion (σgen/σreal mean=1.14). Conclusions: A two-subtype clinical partition, validated longitudinally against DaTscan biomarkers and supported by a converged conditional generative model, provides a defensible imaging-validated stratification framework for PD; the sensitivity grid clarifies the configuration space and strengthens the case for the retained methodology. Replication on independent cohorts (AMP-PD, PDBP, OPDC) is required before clinical translation.