Designing an Automatic Sleep Staging System Using Deep Convolutional Neural Network Fed by Nonlinear Dynamic Transformation

Purpose Studies have shown that sleep significantly affects mental and physical health and quality of life. Sleep staging is one of the most effective diagnostic and therapeutic strategies for sleep disorders. Manual sleep scoring can be a time and cost-consuming process that has a limited level of reliability among raters. This necessitates the use of computer-aided sleep staging. The main objective of the present study is to design an accurate and robust automatic sleep stage scoring system using convolutional neural networks (CNNs) and nonlinear dynamics methods. Methods Since deep learning techniques are effective at classifying images, it is common to convert signals to images first, then feed the images to CNNs. In this study, we propose a new approach for the signal to image transformation, based on the recurrence plot of Polysomnography (PSG) and its frequency characteristics. We evaluated our model using data from 20 subjects of the Sleep-EDF expanded database. Results For five-state classification, our model achieved accuracy, MF1, and Cohen’s Kappa values of 92.5%, 87.1%, and 0.89, respectively. The proposed transformation also significantly improved the detection of the S1 (N1) stage with an F1-score of 0.71. Conclusion The findings of our study demonstrated that a CNN fed by the proposed transformation, which elicits nonlinear characteristics and hidden dynamical patterns in PSG recordings, can enhance performance and improve the efficiency of a sleep staging system.