Two-Dimensional Deep Learning Based Classification of Sleep Stages with Time-Frequency Maps of Single-Lead EEG Segment

Sleep stages classifications are often manually annotated by experts, which are time-consuming, labor-intensive, and have low accuracy. In this work, an automatic classification algorithm of sleep stages by applying the deep convolutional neural network (CNN) to the single-lead electroencephalogram (EEG) was proposed. Specifically, over 84 hours of EEG records were collected from the publically available PhysioNet database, and each of which was segmented continuously into 30-s epochs with labeled annotation of sleep stages, such as W, REM, N1, N2, N3, or N4. Then, each epoch was converted into a two-dimensional (2D) time-frequency map by using the continuous wavelets transform (CWT), where an interesting finding of isolated event was used in the subsequent classification. With a total of 10,000 maps, a 2D thirteen-layer CNN model was developed and used to recognize each map by extracting its deep features automatically. Finally, the proposed 2D CNN based algorithm was evaluated by classifying sleep state into 2, 3, 4, 5 and 6 stages. The obtained results of the model for 6 sleep stages show the accuracy of 82.9%. Moreover, the obtained classification accuracy for 2, 3, 4 and 5 sleep stages were 92.2%, 86.6%, 84.4%, and 83.2%, respectively. Future applications of this classification algorithm would be potentially useful for recognition of sleep stages and diagnosis of sleep related diseases.