Sleep Stage and Apnea Classification from Single-Lead ECG Using Artificial and Spiking Neural Networks

This study presents a comparative analysis of Artificial Neural Network (ANN) and Spiking Neural Network (SNN) models for sleep stage classification and apnea detection using single-lead electrocardiogram (ECG) signals from the ISRUC-Sleep dataset. Both models leverage a deep convolutional recurrent (DCR) architecture, with the SNN variant using spiking neurons in the convolutional layers. Our findings show that the SNN implementation most often surpasses its ANN counterpart. Specifically, our models achieved state-of-the-art performance for binary (Sleep-Wake), three-class (Wake-NREM-REM), and five-class (Wake-N1-N2-N3-REM) classification with the Macro-F1 scores of 0.9451, 0.8161, 0.6843 respectively, surpassing existing methods in all cases. While our models also demonstrated potential in apnea detection, their performance was constrained by the severe class imbalance present in the dataset.