DCM-ML: An Electroencephalography-based Classifier for Early Diagnosis of Schizophrenia Based on Dynamic Connectivity Matrices and Machine Learning Algorithms.

Introduction: Early diagnosis of schizophrenia (SZ) remains challenging due to the subjective nature of clinical assessments and the heterogeneity of symptoms. There is a pressing need for objective, scalable, and non-invasive diagnostic tools to complement traditional methods. This study aimed to propose a machine learning (ML) framework that utilizes dynamic connectivity matrices (DCMs) derived from event-related potentials (ERPs) for SZ classification. Methods: ERP data from 81 participants, including 49 patients with SZ and 32 healthy controls, were sourced from a publicly accessible and anonymized dataset. Granger causality was employed to compute 64×64 directional connectivity matrices, capturing inter-electrode information flow. Feature selection through t-tests identified 2,777 significant connectivity differences (P<0.05), which were subsequently used to train a random forest (RF) classifier. To address class imbalance, balanced training subsets were created. Additionally, the model's robustness was evaluated across varying levels of white Gaussian noise (0% to 45%). Results: The RF classifier demonstrated high diagnostic accuracy (99.24%), sensitivity (98.34%), specificity (99.73%), and an F1-score of 98.91% across 100 iterations, effectively minimizing the risks of overfitting. Its performance remained robust across various train-test splits and substantial noise levels, with an F1-score of 92% even with 45% white Gaussian noise. Feature selection significantly enhanced noise resilience and classification stability. Connectivity analysis revealed that central (Cz, FCz), occipito-parietal (PO3, Oz), and inferior (Iz) regions were key discriminators, indicating disrupted fronto-temporal and sensory integration networks in individuals with SZ. Conclusion: This study highlights the feasibility of ML-driven ERP connectivity analysis as a non-invasive tool for early SZ detection. Achieving near-perfect accuracy, the model demonstrates strong generalizability, interpretability, and clinical scalability, outperforming deep learning counterparts while relying on a minimal, targeted feature set. These findings underscore the diagnostic relevance of fronto-central and occipito-parietal connectivity patterns. While promising as a non-invasive diagnostic adjunct, future validation on larger, demographically diverse cohorts is essential.