EEG-based schizophrenia detection using handcrafted biomarkers and a TOA-optimized hybrid multi-branch
Khosro Rezaee1, Mahshid Dehghanpour2, Maryam Saberi Anari3
1Department of Biomedical Engineering, Meybod University, Meybod, Iran.
Brain Research Bulletin
|April 26, 2026
Summary
This study introduces a novel hybrid framework for schizophrenia diagnosis using electroencephalography (EEG). The advanced model fuses deep learning and traditional biomarkers, achieving high accuracy in subject-level classification and demonstrating strong generalization across different patient cohorts.
Area of Science:
- Neuroscience and Computational Psychiatry
- Biomedical Engineering and Signal Processing
Background:
- Schizophrenia diagnosis relies on clinical assessment, with electroencephalography (EEG) offering a non-invasive tool for analyzing neural dynamics.
- Existing EEG-based computer-aided diagnosis (CAD) methods often use single features and lack subject-level clinical translation.
- There is a need for robust, generalizable EEG analysis frameworks for reliable schizophrenia screening.
Purpose of the Study:
- To develop a hybrid, calibration-aware CAD framework integrating deep learning and handcrafted EEG biomarkers for schizophrenia diagnosis.
- To optimize feature selection and classifier hyperparameters using a novel Treble Opposite Algorithm with feature-search hybridization (TOA-FSH).
- To evaluate the framework's performance and cross-cohort generalization for subject-level classification.
Main Methods:
- Resting-state multi-channel EEG data from two cohorts (MSU-SZ and RepOD-SZ) were preprocessed and converted into continuous wavelet transform (CWT) scalograms.
- A hybrid model fused deep features (EfficientNetV2-S, MobileNetV3-Large, ResNet-50, Swin Transformer) with handcrafted spectral, complexity, and connectivity features.
- The TOA-FSH algorithm selected optimal features and tuned a k-nearest neighbor classifier, followed by probability calibration and subject-level fusion.
Main Results:
- The hybrid framework achieved high window-level (approx. 97.6%) and subject-level (100%) balanced accuracy on the MSU-SZ cohort.
- The model demonstrated strong cross-cohort generalization, achieving similar performance on the unseen RepOD-SZ cohort without retraining.
- Ablation studies confirmed the benefits of ensembling, TOA-FSH feature selection, and hybrid feature fusion over single-backbone approaches.
Conclusions:
- Calibration-aware hybrid modeling using fused deep and handcrafted EEG features provides reliable and transferable subject-level markers for schizophrenia.
- The developed framework shows significant potential for clinician-guided schizophrenia screening and assessment.
- Harmonized preprocessing and evaluation schemes are crucial for achieving robust cross-cohort generalization in EEG-based CAD.