LEEGNet: Lightweight EEG Sleep Stage Classification Network with Knowledge Distillation

Sleep disorders have become increasingly common in recent years, seriously affecting people's quality of life and health. According to the Chinese Sociological Association in 2023, 38.3% of adults in China are diagnosed with sleep disorders. There are many causes of sleep disorders such as stress, e.g., depression, and sleep breathing disorders. Therefore, treating sleep disorders require different approaches based on the different symptoms, but understanding the cause of sleep disorders requires hours of manual labeling by sleep specialists, which is inefficient in the clinical neurology. Although deep learning-based automated sleep stage classification models have achieved clinical accuracy, high complexity and computational resource requirements limit their widespread application. To address this, we propose Lightweight EEG Network (LEEGNet), which introduces knowledge distillation for the first time in the sleep stage classification task. This approach significantly keeps the parameters and complexity to a minimum while maintaining an acceptable level of accuracy, making the model more easily deployed in mobile devices, reducing the time required for manual annotation and improving the speed of diagnosis in clinical neurology. The number of parameters in LEEGNet is 0.9M, which is 2.47 times smaller as compared to the 2.4M of parameters in the state-of-the-art (SOTA) model SleePyCo. Furthermore, LEEGNet runs five times faster on a single-core CPU, and achieves an accuracy of 82.4%, which is only 2.2% lower than the SOTA model.