A CNN-Transformer-ConvLSTM-CRF Hybrid Network for Sleep Stage Classification

Nowadays, many methods have emerged in the research direction of sleep stage classification. Nevertheless, previous researches have solely considered extracting the local features of electroencephalogram (EEG) signals while overlooking the global features and the spatial relationship among these features. More importantly, they disregard the transformation relationship between sleep stages in adjacent epochs. To tackle the problems, we propose a CNN-Transformer-ConvLSTM-CRF hybrid model for sleep stage classification. It is made up of four crucial components: a local-global feature extraction (LGFE) module, a spatio-temporal encoder, an adaptive feature calibration (AFC) module, and a conditional random field (CRF) module. Specifically, multiscale convolutional neural network (MSCNN) and Transformer make up the feature extraction module. MSCNN is capable of extracting EEG signals of different frequencies, and Transformer is used to obtain global features of EEG signals. Then, the STE which is composed of convolutional long short-term memory (ConvLSTM) can capture spatio-temporal relationship among EEG features. Subsequently, the AFC module consisting of a squeeze and excitation (scSE) block enables integrate information of channel features and spatial features. After that, our model outputs the initial classification results. Eventually, the CRF module is capable of learning the transition rules among sleep stages very well to enhance the accuracy of classification. We assess our model on three datasets and it outperforms state-of-the-art methods.