Multivariate connectivity deviations reveal task-evoked and clinical network reorganization

Understanding how cognitive demands and pathology reshape functional connectivity (FC) requires multivariate, region-specific methods. Here, we introduce one-class SVM-based connectome anomaly recognition (OSCAR), which detects condition-related deviations in regional connectivity profiles. OSCAR learns region-to-whole-brain connectivity patterns from a reference state using one-class support vector machines and quantifies target-condition deviations based on outlier proportions. We validated OSCAR across three diverse cognitive tasks, in which it identified regions within networks associated with conflict processing, object-location memory, and lexical learning. In a clinical cohort that included individuals with early psychosis, OSCAR identified relevant regions, including thalamic and basal ganglia areas. Compared with permutation-based multivariate analysis of variance, OSCAR detected additional established task- and disease-relevant parcels, while its task-related findings were at least as close, and often closer, to independent activation findings. OSCAR thus provides an interpretable, region-centered approach for mapping multivariate FC reconfigurations across experimental and clinical conditions.