Structural covariance & graph-learning for the individualized classification of schizophrenia patients

Schizophrenia is considered to be a complex illness with multiple pathways contributing to its clinical phenotype and is argued to be a disorder of disrupted brain network organization. Structural covariance in schizophrenia indicates a system-level brain maturation disruption overlapping with patterns of disturbed functional connectivity. Structural covariance networks are typically computed at the group-level, which prevents individualized inferences. To advance precision medicine of schizophrenia, a reliable mapping of brain system-level heterogeneity at the subject-level is critical. Here, we computed individual ‘Deviation from a Normative Sample’ structural covariance networks (SCN) as networks based on regional gray matter volume (GMV). The nodes in an individual’s SCN represent brain regions of a given parcellation and edge weights quantify the individual’s deviation of pair-wise regional structural covariance from the normative structural covariance of a group of healthy controls. We then assessed their diagnostic value compared to voxel-based and regional GMV for the classification of patients with chronic schizophrenia (N = 71) and healthy controls (N = 74) in a series of support vector machines (SVMs). To assess the richness of information contained in graph-structured data, we trained models on different feature-sets spanning from local edge-weights to global network metrics. All the models performed significantly better than chance with the model trained on the SCNs edge-weights performing the best, reaching a balanced accuracy of 72.3% (sensitivity: 66.2%, specificity: 78.4%). These results highlight the potential of individual SCNs derived from cost-efficient and widely available T1 images. As a primer, we further discuss the utility of more sophisticated graph-learning methods for brain connectivity, such as graph-kernels and graph neural networks.