An explainable hierarchical machine learning framework for schizophrenia biomarker detection based on electroencephalogram signals
Year: 2026
Type: preprint
Abstract: Background Schizophrenia (SZ) is a severe psychiatric disorder that requires objective biomarkers to support clinical diagnosis. Electroencephalography (EEG) provides a non-invasive and cost-effective approach for automated SZ detection; however, many previous studies have been limited by small sample sizes and suboptimal feature selection. This study utilized the ASEEG dataset comprising 101 demographically matched participants (51 SZ and 50 healthy controls). A comprehensive set of resting-state EEG features was extracted under eyes-closed (EC) and eyes-open (EO) conditions. Sequential Forward Selection (SFS) identified the optimal 20-feature subset, and ten machine learning classifiers were evaluated using leave-one-subject-out (LOSO) cross-validation. SHapley Additive exPlanations (SHAP) were employed to interpret model predictions and identify the most influential biomarkers. Results Under the EC condition, Random Forest and XGBoost achieved the highest classification accuracy (81.44%), with Random Forest yielding the best discrimination (AUC = 0.877). Under the EO condition, XGBoost achieved the highest accuracy (78.35%), while MLP attained the highest AUC (0.853). SHAP analysis identified permutation entropy at the frontal midline electrode Fz as the most influential feature. Additional important biomarkers included alpha-band power at the right temporal region (T4), Higuchi fractal dimension, median frequency, and gamma-band mean power at the left temporal region (T3), together with the peak ratio at the right central region (C4). These results demonstrate that combining nonlinear complexity measures with spectral EEG features improves schizophrenia discrimination. Conclusion The proposed framework, integrating SFS-based feature selection with explainable machine learning, achieved robust schizophrenia classification while providing interpretable EEG biomarkers. The findings highlight the complementary value of nonlinear and spectral EEG features and support the potential of explainable EEG-based models as objective tools for schizophrenia diagnosis, warranting further validation for clinical translation.
Source: Research Square
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7204866802