SE-1DCNN-LSTM: A Deep Learning Framework for EEG-Based Automatic Diagnosis of Major Depressive Disorder and Bipolar Disorder

Ziyu Zhao, Hui Shen, Dewen Hu, Kerang Zhang

Abstract: As two typical subtypes of depression, bipolar disorder (BD) is often misdiagnosed as major depressive disorder (MDD) in the early stage. Accurate diagnosis can provide effective treatment for patients. In this paper, we propose a deep learning framework namely SE-1DCNN-LSTM to automatically learn the latent EEG features of the two subtypes. Firstly, a SE block was used as a channel attention module to adaptively learn the weight of each electrode. Subsequently, a 1DCNN-LSTM network was applied to learn discriminative and effective patterns of EEG for MDD and BD. The noteworthy performance of proposed method was verified in 44 MDD and 26 BD patients with 81.10% and 83.16% classification accuracy in epoch-level and subject-level respectively. An extensive investigation of ablation analysis and window size of EEG epoch were conducted. Through visual analysis of electrode weights, we found that the weights of Fp1, Fp2, O1 and O2 electrodes were slightly greater. It demonstrated that the prefrontal lobe and occipital lobe may be possibly important brain regions for MDD and BD recognition. Overall, this study shows the effectiveness of the proposed model in EEG-based automatic diagnosis for MDD and BD.