Neurobiologically Informed AI for Schizophrenia Diagnosis: A Study on Verbal Auditory Hallucinations Using Graph Learning

Early detection of schizophrenia, particularly in individuals exhibiting verbal auditory hallucinations (VAHs), remains a critical challenge in psychiatric research and clinical intervention. Recent advancements in neuroimaging and machine learning have opened new avenues for identifying early biomarkers associated with functional brain network disruptions. In order to increase diagnostic accuracy in differentiating between healthy controls and schizophrenia patients with VAHs, we provide a unique hybrid framework in this study that combines Graph Neural Networks (GNNs) with XGBoost. Using publicly available task-based functional MRI datasets from OpenNeuro, brain graphs were constructed by computing functional connectivity between 116 anatomical regions of interest (ROIs). GNNs were employed to capture topological and spatial relationships between ROIs, while XGBoost exploited both GNN-derived embeddings and auxiliary features–including clinical scores and anatomical measures to perform classification. Our approach achieved a superior accuracy of 94.2% and an AUC-ROC of 0.96, outperforming traditional classifiers such as SVMs, CNNs, and standalone GNN or XGBoost models. SHAP value analysis and graph attention processes, which emphasized important brain regions such the left superior temporal gyrus (STG), anterior cingulate cortex, and Heschl's gyrus locations known to underlie auditory hallucination circuits, guaranteed interpretability. An ablation investigation validated the crucial role of multimodal feature fusion and GNN embeddings. These findings not only validate the predictive strength of the proposed method but also underscore its potential for clinical translation in early-stage psychosis screening. The integration of graph-based learning with ensemble techniques offers a scalable, interpretable, and neurobiologically grounded framework for next-generation schizophrenia diagnostics.