StApneaNet: A Deep Learning-Based Automatic Sleep Stage Adaptive Apnea Detection Network Using Single Channel EEG Signal

Abstract

Sleep apnea is a serious sleep disorder that is characterized by abnormal pauses in breathing. It is related to neural activity that depends on the associated sleep stage of an apnea event. Electroencephalography (EEG) signal is widely used for neural activity analysis and can play a significant role along with sleep stage information in sleep apnea event detection. In this paper, a sleep stage adaptive deep neural network based apnea detection scheme, namely StApneaNet is proposed using single channel EEG signal. The proposed StApneaNet consists of two major blocks: joint model and decision model. In the joint model, multi-band EEG signals are used as input to a multi-kernel CNN block which gives time sequential inter-band related features through causal convolution. A residual squeeze and excitation based channel attention mechanism is then applied to the output feature channels which are further processed through a bi-directional long short term memory (Bi-LSTM) layer along with a temporal attention block. Both apnea prediction and sleep stage prediction are jointly optimized in the joint model that is initially pre-trained and later integrated as a non-trainable block with the trainable decision fusion block inside the decision model for the final apnea event detection. The whole trained network so built is capable of efficiently predicting apnea frames in the testing phase. Extensive experimentation is carried out on three publicly available EEG datasets for subjects-combined and subject-independent classification of apnea events and normal events of an apnea patient. It is found that the proposed StApneaNet offers 91.79% sensitivity, 88.13% specificity, and 89.96% accuracy for a 10-fold subjects-combined cross-validation scheme which are better than those of the existing methods.