Co-Loss Function for Rodents Sleep Stage Scoring Based on Single-Channel EEG

Sleep stage scoring of rodents is important to study the pathological significance of altered sleep. In tradition, sleep experts classify electroencephalogram (EEG) signals into three cardinal sleep stages: Wake, NREM, and REM, which is laborious. Some existing automatic staging methods are over-reliance on domain knowledge and manual labels during extracting features. To solve these problems, we propose a new loss function named Co-Loss to train a contrastive learning model based on single-channel EEG end-to-end. We evaluate Co-Loss in a public rodent sleep dataset from three independent sleep labs. The accuracy is achieved as 88.85%, improving 0.9% compared with Cross-Entropy. This demonstrates, that without changing the model architecture, our loss function can learn features effectively from single-channel EEG with less labeled information and domain knowledge.