Sleep Arousal Detection Using End-to-End Deep Learning Method Based on Multi-Physiological Signals

We propose an end-to-end deep learning method to detect sleep arousals, especially non-apnea sleep arousals, which is the aim of Physionet/CinC Challenge 2018. We use filtered multi-physiological signals as the input of the network without any other hand-crafted features. The network automatically selects the best features to match arousal targets that we want to identify, and outputs the test result. The proposed network architecture is a 35-layer convolutional neural network (CNN) with three parts: a linear spatial filtering with 1 CNN layer, 33-layer Residual Networks (ResNets), and 1 fully connected layer. For the multi-physiological signals provided in the dataset we choose the 6-channel electroencephalography (EEG) and the 3-channel electroencephalography (EMG) signals, since these signals can better represent the characteristics of non-apnea sleep arousals. In the prediction phase, we use a sliding window method to maximize the performance of sleep arousals detection. For the training set, the result of the area under the precision-recall curve (AUPRC) is 0.3173; the area under the receiver operating characteristic curve (AUROC) is 0.8646. For the final test subset, the result of AUPRC is 0.315; AUROC is 0.858.