Optimized 3D CNN with Vision Transformer for Schizophrenia Diagnosis using EEG Signals

This is the authors' abstract. We don't add key points for this paper.

Schizophrenia, a complex and debilitating mental disease, is characterized by disturbances in behaviour, perceptions, and mental processes. Identification at early instances and critical diagnosis are crucial for improving treatment outcomes with regard to enhance the likelihood of patient care and quality. Electroencephalography (EEG) signals serves as a cornerstone in identification of abnormalities associated with schizophrenia. In this manuscript, Optimized 3D CNN with Vision Transformer for Schizophrenia Diagnosis using EEG Signals (3DCNN-ViT-DS-EEG) is proposed. Initially, input data is taken from EEG schizophrenia Dataset. Subsequently, the input data is preprocessing using Multi-window Savitzky-Golay Filter (MSGF) to remove unwanted noise from the input data. As feature extraction plays a major role in medical diagnosis, pre-processed data is input to feature extraction. High-order Time-reassigned Synchro squeezing Transform (HTST) technique is exploited to extract the features like Instantaneous Frequency, Time-Frequency Energy Distribution, Amplitude Modulation and Phase Information .Furthermore, the extracted features are systematically processed through3D CNN with Vision Transformer (3DCNN-ViT),enabling the classification of subjects to ascertain whether the subject is afflicted with schizophrenia or not. Finally, Wolf Bird Optimization (WBO) algorithm is employed to optimize the weight parameter of 3DCNN-ViT classifier, which detects the schizophrenia disease accurately. The proposed 3DCNNViT- DS-EEG is analyzed with performance metrics, like Accuracy, F1-Score, Precision, Computation Time and ROC. The proposed model attains 31.97%, 24.29%, 17.88%, 26.84%, 18.65% and 32.17% higher accuracy,23.80%, 18.64%, 31.79%, 16.24%, 30.89% and 23.16%higher precision and 18.49%, 25.89%, 30.66%, 20.15%, 26.14% and 31.30%lower computational time compared with existing methods respectively.