MEG-based neural biomarkers and AI for early schizophrenia identification

Early schizophrenia detection is vital for timely intervention, and magnetoencephalography (MEG) offers a promising tool for capturing subtle neurophysiological disruptions. However, machine learning (ML) performance in psychiatric diagnosis depends heavily on feature relevance. In this study, we evaluated two distinct feature paradigms: Graph-Theoretical Metrics (GTM; 77 topological measures derived from MEG connectivity); and Expert-guided Neurocognitive and Graph-Theoretical Metrics (ENGTM; GTM integrated with 42 domain-expert-selected clinical/cognitive dimensions, such as MCCB and IQ). MEG data from 52 participants (early-stage schizophrenia vs. healthy controls) were analyzed using six ML classifiers within a nested cross-validation framework. Models trained on GTM features showed moderate performance, with AdaBoost achieving an AUC of 0.77. In contrast, ENGTM-trained models demonstrated significantly superior diagnostic capacity. The Decision Tree model reached 0.96 accuracy (0.97 sensitivity), while Random Forest and SVM achieved near-perfect precision and specificity (0.98). These findings demonstrate that expert-guided feature selection substantially enhances ML performance in MEG-based schizophrenia detection. By bridging clinical psychiatric insight with computational pipelines, this approach improves diagnostic precision and supports the development of reliable, multi-modal neuroimaging biomarkers for clinical use.