Machine learning on multi-modal MRI for early-stage differential diagnosis of Parkinson's disease and Parkinson-plus syndromes

The early differential diagnosis between Parkinson's disease (PD) and Parkinson-plus disease (PPS) is often challenging due to overlapping clinical expressions. This study aims to assess the feasibility of employing multi-modal MRI combined with machine learning for the differentiation of these conditions. The T1-weighted, diffusion-weighted and susceptibility-weighted images from healthy controls (n=67), PD patients (n=75) and PPS patients (n=29) were collected. For HC versus PD, the optimal performance was achieved by the integration of QSM with T1 or DTI with an accuracy of 86%, recall of 80%, precision of 100% and f1 score of 89%. For HC versus PS, DTI performed best with an accuracy of 73%, recall of 100%, precision of 57% and f1 score of 73%. For PD versus PS, the combinations involving QSM exhibited the optimal performance with an accuracy of 83%, recall of 100%, precision of 50% and f1 score of 67%. Our findings suggest that multi-modal MRI combined with machine learning holds significant potential for the early differential diagnosis of PD and PPS.