Improving Sleep Stage Classification Performance by Single-Channel EEG Data Augmentation via Spectral Band Blending

Classification of sleep stage for neurophysiological signals acquired during polysomnography is an important task in the diagnosis of sleep disorders and sleep research. However, current manual sleep stage labeling by trained experts involves heavy labour intensity. Furthermore, previous models regarding the automation of sleep stage classification using machine learning are highly susceptible to class imbalance problem, which is prevalent in polysomnography recordings. This study proposes an automated sleep stage classification method based on data augmentation via spectral band blending to address class imbalance problem, and thereby improving the performance of machine learning models for sleep stage classification. In this endeavour, EEG recordings from 153 healthy subjects were utilized to develop the EEGNet-BiLSTM as the baseline model. When compared the performance of sleep stage classification between EEGNet-BiLSTM with and without data augmentation, EEGNet-BiLSTM without data augmentation yielded approximately 82% accuracy and 0.70 kappa-value, whereas the proposed method showed 87% accuracy and 0.73 kappa-value. Our proposed method, EEGNet-BiLSTM with data augmentation, is superior in terms of accuracy and consistency compared to the conventional baseline model.