SPSNet: A spiking neural network with relation graphs for sleep stage classification based on polysomnography

Abstract

Sleep is crucial to human health, and in recent years, automatic sleep stage classification based on polysomnography(PSG) has become a hot topic in sleep science research. With the rapid development of artificial intelligence technology, especially the wide application of deep learning methods, the research on automatic sleep stage classification has made significant progress. However, existing methods mainly focus on time–frequency feature extraction and channel selection of signals, often ignoring the deep impact of biological mechanisms such as neuronal impulses on sleep stage classification. To this end, we propose a deep learning model called SPSNet, which innovatively introduces the impulse mechanism of a spiking neural network(SNN) and the structure of relational graph based on the transformation of the Watts–Strogatz(WS) small-world network into the epoch-level multi-channel sleep feature fusion process. This design not only achieves efficient sparse computation through SNN, but also enhances the interaction between neurons and improves the overall model performance through the relational graph structure. Experimental results on three public datasets(UCD, SleepEDF-78, HMC) show that SPSNet significantly improves the classification performance while effectively reducing network complexity and energy consumption compared to the baseline model approach, the accuracy(ACC) on the three datasets were 0.772, 0.807, and 0.775, the F1-score(MF1) were 0.761, 0.758, and 0.756, and the Cohen’s Kappa($\kappa$) was 0.703, 0.739, and 0.706, representing improvements of 0.3% to 1.8% over the respective best baseline models. Overall, our work provides a new way of thinking for automatic sleep stage classification that combines spiking neural networks with relational graph structures.