Deep Learning Automatic Sleep Staging Method Based on Multidimensional Sleep Data

Abstract

Polysomnography (PSG) is commonly used to diagnose sleep disorders. However, manual sleep staging is a time-consuming task due to high human effort and technical thresholds, and it involves certain subjective factors. To improve the efficiency of sleep staging, this paper proposes a deep learning automatic sleep staging method based on multidimensional sleep data, the MCSSN model. This model uses continuous wavelet transform (CWT) to first extract sleep features, followed by data enhancement of time-frequency features using the SMOTE algorithm. Finally, the CNN-BiLSTM model is utilized for training, which involves employing a two-step training algorithm to learn sleep features and the transition rules of sleep states. The evaluation results show that the model achieved an excellent overall accuracy of 85.87% on a sample of 104 subjects, with a macro average score ( $F_{1M}$ ) of 0.82 and a Kappa of 0.80. The model achieved excellent classification performance for N1 stage detection, with an F1 of 0.64, outperforming the other models. The experimental results show that the SMOTE oversampling algorithm plays an active role in detecting sleep stages, especially the N1 stage, which is difficult to identify. In addition, learning sleep transition rules through a two-step training algorithm helps to improve the performance, and the MCSSN algorithm provides a reference for the automatic design of subsequent sleep classification networks.

Keywords

• Polysomnography
• sleep stages
• continuous wavelet transform
• SMOTE
• two-step training algorithm