Abstract Schizophrenia spectrum disorders (SSD) are marked by widespread structural brain abnormalities. Neuroanatomical normative modeling (NM) can quantify person-specific deviations from healthy variability, yet it remains unknown whether pre-trained, large-scale NM features support site-held-out classification and mechanistic brain–behavior mapping in SSD. Here, we applied a publicly available PCNtoolkit model (trained on ~57,000 healthy controls from 82 sites) to six independent cohorts (N = 831) to derive individual deviations in cortical thickness (CT) and subcortical volumes from T1-weighted MRI. Employing a random forest classifier with leave-site-out cross-validation, we achieved a balanced accuracy of 65%, which underscores the inherent complexity of SSD. Feature importance analysis identified total gray matter volume, mean CT, and CT changes in limbic and sensorimotor regions as key predictive features. Relative to healthy controls, SSD participants showed a higher burden of extreme negative deviations, which related to reduced attention and processing speed and to elevated neurological soft signs (NSS). Finally, canonical correlation analysis revealed a robust multivariate relationship linking structural deviationsparticularly CT changes in limbic and sensorimotor cortices, to cognition and NSS. Together, these results demonstrate that NM features transferred from a large external reference can generalize across sites and elucidate clinically relevant brain–behavior associations in SSD, supporting the integration of multimodal, large-scale datasets to advance biomarker discovery and inform earlier, more targeted interventions.
Volkmer et al. (Fri,) studied this question.