SleepExpertNet: high-performance and class-balanced deep learning approach inspired from the expert neurologists for sleep stage classification

Sleep stage classification is crucial in diagnosing sleep disorders and monitoring treatment effectiveness, yet it is inconvenient, requiring many electrodes and labor-intensive assessments. Single electroencephalogram (EEG)-based deep learning approaches were proposed to address this issue; however, existing studies have employed deep learning models that are devised for imaging and natural language processing for automated sleep stage classification, which limits their use for capturing sleep patterns or changes over time in EEG signals. This study proposes SleepExpertNet, which is inspired by actual clinical guidelines for scoring sleep stages. The model consists of a representation learning part, which learns the physiological characteristics of EEG sleep signals, and a temporal context modeling part, which learns long- and short-term temporal context information similar to a sleep expert. SleepExpertNet guarantees the highest level of accuracy (accuracy: 90.8%, macro-f1: 86.7) on the expanded Sleep-EDF dataset, surpassing the existing state-of-the-art methods, despite utilizing only single-channel EEG signals. Furthermore, the class imbalance problem, which has been a major obstacle in sleep stage classification research, was also addressed. The proposed model is expected to reduce the workload on clinicians and support diagnostic decision-making, enabling more accurate sleep stage classification.