EEG-based biomarkers for psychosis: Comparative performance of support vector machines and deep neural networks

Abstract

Electroencephalography (EEG) provides a widely accessible window into neural abnormalities in psychosis-spectrum disorders, yet the comparative utility of classical machine learning and deep learning under limited-sample conditions remains uncertain. This study compared support vector machines (SVM) and deep neural networks (DNN) in classifying psychosis based on task-related EEG collected during the Ultimatum Game. Data from 43 participants (19 patients, 24 controls) were preprocessed and used to extract spectral features, nonlinear dynamics (Hjorth parameters, entropy, fractal dimension), and functional connectivity measures. Classification was performed using 5-fold subject-wise cross-validation, and feature importance was evaluated via permutation and SHapley Additive exPlanations (SHAP) analyses. Support Vector Machine (SVM) achieved superior performance (Accuracy = 89.9%, AUC = 0.959) relative to DNN (Accuracy = 78.1%, AUC = 0.879). Nonlinear features, particularly Hjorth complexity and activity, together with delta/theta power, were the strongest contributors to discrimination. These findings indicate that, in small but feature-rich EEG datasets, classical machine learning provides more stable and interpretable performance than deep learning. Nonlinear and low-frequency indices emerge as promising candidate biomarkers in psychosis-spectrum disorders, with potential applications in early screening, treatment monitoring, and individualized neurofeedback. Larger multisite datasets will be essential to validate their generalizability and clinical utility.