Spectral and Graph Based Deep Learning Models for Schizophrenia Detection

Deep learning has intervened in the field of neuroscience by inspecting activities in the brain through Electroencephalogram (EEG) signals. This research investigates the application of two unconventional preprocessing approaches to EEG signals, aiming to enhance the classification of schizophrenic adolescents from healthy individuals. To analyze the EEG signals, the researchers used three different mathematical techniques called Fourier Transform (FFT), Short-Time Fourier Transform (STFT) and Laplace Transform. These techniques helped to extract important features from the signals, or to identify patterns in the brain activity that could be related to schizophrenia. After extracting the features, the researchers then used them as input for different deep learning models to detect schizophrenia. Deep learning models like VGG-16, Variational AutoEncoders, ChronoNet and GraphCNN were used. It was found that the STFT processed signals on VGG-16 gave the best results. Certain challenges like variation in data, the need for the large annotated datasets along with the interpretability of deep learning models were encountered. To address these issues, this study deployed preprocessing techniques and the domain knowledge has been integrated to enhance the model robustness and generalizability.