Diagnosis of Schizophrenia in EEG Signals Using dDTF Effective Connectivity and New PreTrained CNN and Transformer Models

Schizophrenia (SZ) is a multifaceted mental disorder that typically emerges in early adulthood, characterized by a spectrum of physiological and cognitive deficits. Electroencephalography (EEG) recordings are pivotal in SZ diagnosis, necessitating the expertise of specialist doctors and psychologists. However, the analysis of EEG signals is labor-intensive and susceptible to human error. This study introduces a deep learning (DL) pipeline for the early detection of SZ using EEG signals. The pipeline includes stages of dataset selection, preprocessing, feature extraction, and classification. For this study, the RepOD dataset, consisting of EEG recordings from 14 subjects with SZ and healthy controls (HC), was utilized. The preprocessing phase involves normalizing and segmenting the EEG data. Subsequently, the EEG signals are divided into various sub-bands via Discrete Wavelet Transform (DWT), and effective connectivity matrices are derived using the directed Directed Transfer Function (dDTF) technique. Following this, state-of-the-art pretrained DL models based on CNNs and transformers are applied to extract features and classify the 2D dDTF images obtained from different EEG sub-bands. Notably, the ConvNext-Tiny architecture demonstrated superior performance, achieving an accuracy of \(96 \%\) in the beta sub-band. Furthermore, this model surpassed the performance of other DL models in terms of accuracy across additional EEG sub-bands.