OBJECTIVES: To develop a network-based machine learning framework for classifying psychosis using EEG connectivity features and to identify stable, reproducible candidate markers through a stability-driven approach. METHODS: This study was designed as a cross-sectional analytical secondary data analysis using a publicly available EEG dataset comprising 43 participants (19 psychosis, 24 controls). Functional connectivity measures (PLV and coherence) and graph-theoretical network features were extracted across frequency bands. A nested cross-validation framework with subject-wise splitting was applied. Feature importance was evaluated using permutation importance and SHAP, and stability scores were computed across folds to identify robust features. Classification was performed using support vector machine (SVM) and random forest (RF) models. RESULTS: The SVM model achieved superior performance (accuracy: 91.2%, AUC: 0.967). Stability analysis identified theta-band connectivity features, particularly fronto-parietal PLV and global efficiency, as the most reliable and discriminative. SHAP analysis confirmed their consistent contribution across subjects. However, these findings should be interpreted as exploratory given the relatively small sample size and the use of a single dataset. CONCLUSION: Stable connectivity-derived network features provide interpretable and robust candidate EEG markers for psychosis classification. The proposed framework enhances reproducibility and supports the potential of EEG-based tools for clinical screening, while emphasizing the need for external validation.
Naeim et al. (2026) studied this question.