The study presents a machine-learning pipeline that produces a patient-specific heterogeneity score (HET) from multimodal MRI with SHAP explainability, distinguishing MSA subtypes from PD, identifying key affected regions (olivopontocerebellar and striatonigral areas and widespread white matter),…

Delivers a quantitative, explainable imaging biomarker useful for diagnosis, patient stratification, and sensitive longitudinal monitoring—tools that can improve clinical trial selection and outcome measurement and thereby indirectly accelerate therapeutic development despite not revealing…

UNLABELLED: Disease progression in multiple system atrophy (MSA) and Parkinson’s disease (PD) shows marked patient-to-patient heterogeneity. We hypothesize that machine learning methods applied to multimodal MRI data would aid in optimally identifying critical brain regions impacted in each patient, improve disease differentiation and longitudinal tracking. Using structural and diffusion MRI of MSA (cerebellar and parkinsonian subtypes), PD, and normal participants, we trained binary classifiers and utilized Shapley Additive exPlanations (SHAP) to quantify feature contributions to derive heterogeneity scores (HET). HET outperformed commonly available imaging tools when differentiating between MSA and PD, strongly correlated with clinical markers, and sensitively tracked longitudinal disease progression. HET correctly identified olivopontocerebellar atrophy and striatonigral degeneration as important for disease identification, shed light on the spatio-temporal disease progression, and identified widespread white matter involvement in MSA. Our machine learning approach quantifies MSA and PD heterogeneity and provides a patient-specific measure for precise disease quantification and longitudinal tracking. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1038/s41598-026-45949-5.