Structural parameters are superior to eigenvector centrality in detecting progressive supranuclear palsy with machine learning & multimodal MRI.

Progressive supranuclear palsy (PSP) is an atypical Parkinsonian syndrome characterized initially by falls and eye movement impairment. This multimodal imaging study aimed at eliciting structural and functional disease-specific brain alterations. T1-weighted and resting-state functional MRI were applied in multi-centric cohorts of PSP and matched healthy controls. Midbrain, cerebellum, and cerebellar peduncles showed severely low gray/white matter volume, whereas thinner cortical gray matter was observed in cingulate cortex, medial and temporal gyri, and insula. Eigenvector centrality analyses revealed regionally specific alterations. Multivariate pattern recognition classified patients correctly based on gray and white matter segmentations with up to 98 % accuracy. Highest accuracies were obtained when restricting feature selection to the midbrain. Eigenvector centrality indices yielded an accuracy around 70 % in this comparison; however, this result did not reach significance. In sum, the study reveals multimodal, widespread brain changes in addition to the well-known midbrain atrophy in PSP. Alterations in brain structure seem to be superior to eigenvector centrality parameters, in particular for prediction with machine learning approaches.