SimpleKANSleepNet: a Kolmogorov–Arnold network based sleep stage classification method

A novel Kolmogorov-Arnold Network (KAN) based machine learning model is proposed for the automatic sleep stage classification task. The redefined architecture of the Multilayer Perceptron (MLP) aims to build a more flexible model by using learnable activation functions. In this study, an effective KAN model named SimpleKANSleepNet is evaluated on two different datasets with temporal features and frequency features extracted from electroencephalography (EEG), electromyogram (EMG), electrooculogram (EOG), and electrocardiogram (ECG) signals through a dual-stream convolutional neural network (CNN). Compared with existing CNN-based methods and graph convolutional networks (GCNs), the proposed model achieves an overall classification accuracy, F1-score, and Cohen's kappa on the ISRUC-S1 and the Sleep-EDF-153 datasets of 0.812, 0.793, 0.757, 0.928, 0.929, and 0.910, respectively, which demonstrates its competitive classification performance and generality. Moreover, several data balancing methods are tested on Sleep-EDF-153 to further evaluate the potential for achieving the best results. Finally, the factors that may affect the classification ability are tested on the ISRUC-S1 dataset.