Leveraging a novel deep learning model with residual blocks for accurate bipolar disorder diagnosis from EEG data

The diagnosis of bipolar disorder (BD) in clinical psychiatry is very challenging: it requires subjective assessments and self-reported symptoms. It makes timely intervention difficult and results in different outcomes of treatment. New developments in neuroimaging and signal processing have created opportunities for improved psychiatric diagnosis. The deep learning architecture introduced in this research for EEG classification of BD is based on residual block, LSTM, and attention. We present a deep neural network that operates on multiple channels and treats the task of EEG signal classification as an end-to end problem that does not require the use of any feature extraction methods. Convolutional layers are used in the model to capture some of the most root features and the use of LSTM and attention mechanism enhances the generality of the network further. We show that our network can be validated for BD detection on two different EEG datasets with accuracy of 97.10% and 96.02% respectively. It is experimental findings that show that our model has high recognition rate and robust generalization capability relative to previous techniques.