PriorSleepNet: Enhancing Sleep Stage Classification Through Prior Knowledge Integration

The basis for measuring sleep quality is the time distribution across different sleep stages. High-quality sleep should include sufficient sleep duration and a reasonable distribution of sleep stages. Therefore, sleep staging is a key indicator in sleep medicine. This paper introduces a novel sleep staging model, PriorSleepNet, which synergistically combines feature extraction, sequence modeling, and prior knowledge integration to achieve robust and physiologically consistent sleep staging classification. We use ResNet-34 to extract features from each 30-second sleep epoch, and then introduce a transformer structure as a sequence network to model the sequential information between epochs throughout the night. Unlike previous work that split the entire night's data into short sequences for processing, we pre-analyzed the distribution of sleep states as prior knowledge to address the issue of segmentation potentially losing the guiding role of preceding epochs. This allows us to use earlier sleep epochs to assist in determining the sleep stage at the beginning of the segments. Additionally, by modeling the complete sequence, we consider the dynamic changes and distribution of sleep states over time. PriorSleepNet collects labels for the entire night's sleep data, obtaining the statistical distribution of sleep stage counts at various times. This quantifiable prior knowledge of sleep patterns is introduced into the model to correct the model's erroneous staging results, thereby better reflecting biological sleep patterns. Experiment results show that the PriorSleepNet model has achieved excellent performance on the sleep dataset. Compared to existing methods, its staging accuracy has significantly improved, demonstrating the model's effectiveness and superiority in sleep staging tasks.