Abstract
Schizophrenia (SCH) is a chronic and severe mental disorder that leads to significant cognitive and neurophysiological impairments, affecting daily life. Early diagnosis remains challenging as it relies on the manifestation of symptoms that develop over time. Electroencephalography (EEG), which measures brain activity, provides a promising avenue for early detection. In this study, two EEG datasets-the Mental Health Research Center (MHRC) and the Repository for Open Data (RepOD)-were employed to detect SCH. EEG signals were segmented into 8-second durations and decomposed using Variational Mode Decomposition (VMD) into 10 Intrinsic Mode Functions (IMFs). Multi-domain features extracted from IMFs were classified using nine machine learning (ML) and seven optimized ML (OML) classifiers. The proposed method achieved an accuracy (Ac) of 96.7% for the MHRC dataset using the Optimizable KNN classifier and 99.0% for the RepOD dataset using the Optimizable Ensemble classifier. To prevent data leakage, a strict subject-wise Leave-One-Out Cross-Validation (LOOCV) strategy was employed. Lobe-wise analysis showed that the frontal lobe achieved accuracies of 91.2% for MHRC using the Optimizable Ensemble and 99.4% for RepOD using the Optimizable Neural Network, with the temporal lobe also showing strong discriminative power. These findings align with established evidence of frontal-temporal dysconnectivity in SCH. Overall, the proposed VMD + OML framework offers a computationally efficient and clinically interpretable solution for early SCH detection using EEG signals.