Machine learning-based morphological brain analysis in schizophrenia and unaffected siblings: a multisite study of potential risk markers.

Background and hypothesis: Assessing schizophrenia risk factors is crucial for developing early preventive interventions. We hypothesized that unaffected siblings, who share high genetic risk, exhibit neuroanatomical signatures similar to affected patients, potentially reflecting early pathogenic processes. Study design: To overcome single-center limitations, we analyzed 1,018 participants from five independent, public databases. Brain MRIs were standardized via voxel-based morphometry, and covariate-adjusted z-scores were calculated for regional volumes. An ensemble support vector machine (SVM) approach, incorporating multiple models to ensure robustness, was employed to extract a multidimensional brain signature, from which a schizophrenia-like score (SPS) was derived. Results: The ensemble SVM achieved high classification performance (AUC = 0.99861). Across all databases, patients exhibited consistent volume reductions in frontal, temporal, insular, and thalamic regions, alongside globus pallidus enlargement. Notably, unaffected siblings were 3.8 times more likely to show brain morphological similarities to patients than were healthy controls. Furthermore, we identified a novel imaging phenotype in siblings: increased ventral striatal volume, which positively correlated with the SPS. This feature, absent in established schizophrenia, suggests a potential compensatory mechanism or a transient developmental marker of risk. Conclusion: Applying machine learning to large-scale, multi-site neuroimaging data effectively identifies structural endophenotypes. Our findings highlight unique structural characteristics, specifically the enlarged ventral striatum, as a critical biological metric for identifying high-risk individuals before clinical onset.