Schizophrenia classification from multichannel EEG using EEGNet with SHAP and Grad-CAM

Schizophrenia affects cognition, perception and neural information processing. Electroencephalography (EEG) provides a non-invasive and relatively low-cost way to study these changes. This study evaluates two deep learning models for separating paranoid schizophrenia patients from healthy controls using the public RepOD resting-state EEG dataset, which contains 14 patients and 14 healthy controls recorded from 19 electrodes at 250 Hz.

To prevent participant level data leakage, all final testing kept participants separate using Leave-One-Subject-Out (LOSO) and Stratified Group 5-Fold validation. The three-dimensional Convolutional Neural Network (3D-CNN) used Short-Time Fourier Transform representations (STFT), while EEGNet used the channel-time EEG epochs directly. At the reference settings used for the final controlled analysis, EEGNet gave an accuracy of 0.750, F1-score of 0.767, Cohen’s kappa of 0.497 and AUC of 0.797 under LOSO validation.

Under Stratified Group 5-Fold validation, the corresponding values were 0.679, 0.695, 0.357 and 0.757. EEGNet showed stronger performance than the 3D-CNN, particularly under Stratified Group 5-Fold validation, where the 3D-CNN achieved a kappa of 0.011. Participant-based bootstrap confidence intervals quantified the uncertainty associated with the 28-participant cohort.

Additional tests of learning rate (LR) and STFT resolution confirmed that the 3D-CNN was sensitive to model settings and fold composition. A separate ICA cleaning test showed that removing likely artifacts did not correct the weak 3D-CNN stratified result, while EEGNet performance improved after cleaning. SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM) were used to examine which EEG channels and signal regions influenced the model decisions.

The XAI results were clearer for EEGNet, while SHAP and Grad-CAM assigned different rankings to individual electrodes. These findings provide a basis for further validation using larger and independent datasets. Open access funding provided by Vellore Institute of Technology.

This research received no external funding. School of Electronics Engineering (SENSE), Vellore Institute of Technology, Chennai, India The authors declare no competing interests. This study is a secondary analysis of the publicly available RepOD Paranoid Schizophrenia EEG dataset (Olejarczyk & Jernajczyk, 2017; doi: https://doi.org/10.18150/repod.0107441).