Improving the diagnosis of first-episode schizophrenia from magnetic resonance imaging using machine learning

Background: Early diagnosis of schizophrenia could improve the outcomes and limit the negative effects of untreated illness. Although participants with schizophrenia show structural/functional alterations on the group level, these findings have a limited diagnostic utility. Novel methods of MRI analyses, such as machine learning (ML), may help bring neuroimaging from bench to the bedside. Here, we used ML to differentiate participants with a first episode of schizophrenia-spectrum disorder (FES) from healthy controls (HC) based on neuroimaging data and compared the diagnostic utility of such approach with the utility of between group comparisons using classical statistical methods. Method: Firstly, we performed a classical fMRI experiment in FES using a self/other- agency task (SA/OA) and compared FES (N=35) versus controls (N=35) using conventional statistics. We than classified FES and healthy controls (HC) using linear kernel support vector machine (SVM) from the resting-state functional connectivity (rsFC) and fractional anisotropy (FA) in 63/63 and 77/77 age- and sex-matched FES and HC participants. We also investigated the between-group differences in rsFC and FA using classical between-group comparisons. Results: FES group exhibited a decreased activation during the emergent SA experience..