Deep learning based sleep stage classification

Deep learning based sleep stage classification

Sleep plays an essential role in humans’ life and sleep stage classification is the first step for sleep research and sleep disorder diagnosis. This research aims to classify sleep stages automatically using deep learning methods, which can help clinicians identify sleep problems. Polysomnograms (PSGs), including electroencephalography (EEG), electromyogram (EMG), electrocardiogram (ECG), and electrooculogram (EOG), are signals collected through placing electrodes on the scalp cross different locations, which are powerful tools for sleep stages classification and sleep disorders identification. To classify sleep stages more effectively and efficiently, three models were developed in this research, namely the jumping knowledge spatial-temporal graph convolutional network (JK-STGCN) model, the 3DSleepNet model, and the MixSleepNet model. For all the three models, key features are extracted from multi-channel signals to aggregate spatial and temporal information. For the JK-STGCN model, the connections among different bio-signal channels from the identical epochs and their neighbouring epochs can be obtained through two adaptive adjacency matrices learning methods. A jumping knowledge spatial-temporal graph convolution module helps this model to extract spatial features from the graph convolutions efficiently and temporal features are extracted from its common standard convolutions to learn the transition rules among sleep stages. For the 3DSleepNet model, the intrinsic connections among different bio-signals and different frequency bands in time series and time-frequency are learned by the 3D convolutional layers, while the frequency relations are learned by the 2D convolutional layers. The partial dot-product attention layers help this model find the most important channels and frequency bands in different sleep stages. A long short-term memory unit is added to learn the transition rules among neighbouring epochs. For the MixSleepNet model, the 3D convolution branch can explore the correlations between multi-channel signals and multi-band waves in each channel in the time series, while the graph convolution branch can explore the connections between each channel and each frequency band. The experiments on ISRUC-S3 and ISRUC-S1 demonstrate that the JK-STGCN model outperforms the 1D-CNNs, 2D-CNNs, U2-Net, and other GCN models on these two subsets, while the 3D-CNN model achieves similar performances with a faster speed, and the MixSleepNet model achieves better performances based on the second expert’s label with a notable improvement of sleep stage 1 precision.

File reproduced in accordance with the copyright policy of the publisher/author/creator.