EEG-Based Depression Recognition Using Intrinsic Time-scale Decomposition and Temporal Convolution Network

Abstract

The diagnosis and treatment of depression is very important since it brings a heavy burden to family and society. Because of the high sensitivity, relatively low cost, and convenient recording, electroencephalogram (EEG) has become an important tool for monitoring brain activity and is gradually being used in the auxiliary diagnosis of mental diseases. EEG signals are typically non-linear and non-stationary. Therefore, they are suitable to be dealt with by time-frequency analysis technique. In this paper, we propose a strategy that combines the time-frequency analysis technique and temporal convolution network for depression recognition. Firstly, we use the method of intrinsic time-scale decomposition to decompose each EEG recording to several components. And secondly, some statistical indices are calculated from the instantaneous amplitudes and instantaneous frequencies of these components to form the feature vectors. Thirdly, an improved temporal convolution network (TCN) is used to detect the depression from normal controls. Temporal convolution network is not only suitable for the sequence model, but also retains the parallel computing characteristics of the convolutional neural network. To improve the model performance, we further modify the original softmax loss of TCN as L-softmax. Experiments show the effectiveness of the proposed model. Furthermore, we find that the depressive patients and normal controls shows different patterns through functional connectivity analysis. Our analysis results can be used as an auxiliary tool to help psychiatrists diagnose patients with depression.