Automated classification of neurodegenerative parkinsonian syndromes using multimodal magnetic resonance imaging in a clinical setting

Background: Several studies have shown that machine learning algorithms using MRI data can accurately discriminate parkinsonian syndromes. Validation under clinical conditions is missing. Objectives: To evaluate the accuracy for the categorization of parkinsonian syndromes of a machine learning algorithm trained with a research cohort and tested on an independent clinical replication cohort. Methods: 361 subjects, including 94 healthy controls, 139 patients with PD, 60 with PSP with Richardson9s syndrome, 41 with MSA of the parkinsonian variant (MSA-P) and 27 with MSA of the cerebellar variant (MSA-P), were recruited. They were divided into a training cohort (n=179) scanned in a research environment, and a replication cohort (n=182), scanned in clinical conditions on different MRI systems. Volumes and DTI metrics in 13 brain regions were used as input for a supervised machine learning algorithm. Results: High accuracy was achieved using volumetry in the classification of PD versus PSP, PD versus MSA-P, PD versus MSA-C, PD versus atypical parkinsonian syndromes and PSP versus MSA-C in both cohorts, although slightly lower in the replication cohort (balanced accuracy: 0.800 to 0.915 in the training cohort; 0.741 to 0.928 in the replication cohort). Performance was lower in the classification of PSP versus MSA-P and MSA-P versus MSA-C. When adding DTI metrics, the performance tended to increase in the training cohort, but not in the replication cohort. Conclusions: A machine learning approach based on volumetric and DTI data can accurately classify subjects with early-stage parkinsonism, scanned on different MRI systems, in the setting of their clinical workup.