Automated differentiation for parkinsonism using clinical structural MRI: A multicenter study

Multimodal research-grade magnetic resonance imaging (MRI) scans show potential for differentiating Parkinson’s disease (PD) from multiple system atrophy (MSA), but clinical translation is limited by cost, acquisition time, and processing burden. Here, we developed a T1-weighted structural MRI-based machine learning model to differentiate Parkinsonism and generalized it to routine anisotropic 2D clinical MRI. Using a multicenter cohort of 969 participants, we employed a hierarchical discovery–verification–translation framework. While a multimodal benchmark model achieved an area under the receiver operating characteristic curve (AUROC) of 0.98, our streamlined unimodal T1-Top20 model achieved an AUROC of 0.93 for multiclass discrimination (healthy controls, PD, MSA). For PD vs MSA, the frozen T1-Top20 model achieved a pooled AUROC of 0.88 in the external research-grade cohort ( n = 227) and identified 19 of 25 who phenoconverted from idiopathic rapid eye movement sleep behavior disorder (iRBD) to PD. Importantly, the generalized model achieved a pooled AUROC of 0.84 in the clinical-MRI cohort (2D T1, n = 303), supporting its potential for clinical translation.