Detection of Schizophrenia Using Electroencephalography Signals with a Hybrid EEGNet and Transformer Model

Abstract

The diagnosis of schizophrenia has historically been based on clinical opinion, a method which has been observed to lend itself to late diagnosis or non-standard outcomes. EEG offers an affordable, non-invasive window into brain dynamics that may support the development of objective biomarkers for psychiatric disorders such as schizophrenia. This paper proposes a novel hybrid deep learning framework which combines EEGNet's spatial-temporal processing capability with the long-range sequential modelling capability of Transformer encoders. Preprocessing is done using MNE-Python, including bandpass filtering and ICA-based artefact removal to eliminate physiological artefacts. The model is tested on normalised electroencephalography signals, segmented into 2-second epochs, with a test accuracy of 99.72%, precision of 99.73%, recall of 99.71%, and an F1-score of $\text{9 9. 7 2 \%}$. The proposed framework demonstrates strong performance compared to baseline methods and provides a foundation for further research on explainable artificial intelligence and real-time electroencephalography analysis.