GCN-LSTM: A hybrid graph convolutional network model for schizophrenia classification

Schizophrenia is a complex mental disorder that influences one’s perceptions, thought processes, social behavior, emotional responses, etc. Electroencephalography is a non-invasive brain imaging technique that measures the brain’s electrical activity. The EEG signals are used to study and analyze the human brain. Graphs have always been one of the best ways to represent information. With the inspiration from graphs, in this paper, we developed a novel GCN-LSTM model, a graph-based hybrid deep learning model for classifying schizophrenia from Healthy Control. We used the Institute of Psychiatry and Neurology in Warsaw, Poland dataset to experiment with the developed models; Raw EEG signals were pre-processed and divided into segments of 5-sec and 8-sec. We extracted 14 different features from these epochs, 7 each from the time and frequency domains. After feature extraction, we constructed the graphs out of epochs of 5-sec and 8-sec, where EEG electrodes are considered as nodes and how signal flows between EEG channels as edges. These graphs were fed to the developed GCN-LSTM model for the classification. We also used different seeds and 5-fold cross-validation to avoid overfitting. We conducted several experiments and achieved average accuracy across all seeds as 99.25 ± 0.24 for the GCN-LSTM model with 8-sec epoch data, also Precision of 99.28 ± 0.22 %, F1 score of 99.24 ± 0.24 %, Specificity of 98.73 ± 0.64; Sensitivity of 99.67 ± 0.28 and AUC of 99.20 ± 0.27. We used t-test and one-way ANOVA to study the statistical significance of the extracted features. We found zero crossing rate, mobility (Hjorth parameter), peak frequency, and gamma band.