SleepNet-Lite: A Novel Lightweight Convolutional Neural Network for Single-Channel EEG-Based Sleep Staging

Abstract

Sleep is indispensable for human survival, and automatic sleep staging is crucial to evaluating sleep quality. Deep learning has become popular in recent years as a way to increase the accuracy of sleep staging. However, most methods have a large number of parameters and big computational burden that are not suitable for mobile deployment. To this end, we propose a novel lightweight model named SleepNet-Lite for sleep staging. First, a multiscale convolutional block (MSCB) is designed to extract multiscale electroencephalogram features. Second, an inverted residual block is used to fuse the multiscale features. Channel shuffle and channel split operation are inserted into the MSCB and residual block, realizing the interactive information flow to improve performance. In addition, we replace the standard convolution with depthwise separable convolution to increase representational efficiency and reduce the number of parameters. Then, the global average pooling layer is utilized before fully connected layer to further reduce the parameters and avoid overfitting. On account of class imbalance, we use the logit-adjustment loss function to adaptively pay attention to each stage in the training process without increasing the computational burden. We show that the SleepNet-Lite surpasses the state-of-the-art approaches, in overall accuracy and Cohen's Kappa on two public datasets. We achieve superior performance with fewer parameters of 41.67 K on the Seep-EDF dataset and 42.44 K on the MASS-SS3 dataset, providing a promising deployment solution for mobile platforms and wearables.