Investigating the use of uni-directional and bi-directional long short-term memory models for automatic sleep stage scoring

In this paper, a study is conducted to investigate the use of a Long Short-Term Memory (LSTM) learning system in automatic sleep stage scoring. The developed algorithm will automatically learn to classify sleep stages from any acquired sleep signals data-set. This allows to resolve the difficulties that are facing experts in manual sleep stage scoring. A total of 39 Polysomnogram (PSG) recordings acquired from the online PhysioNet Sleep-EDF database are used in this study. The PSG recordings are chosen to be from the EEG Fpz-Cz signals only. The database comes with annotation files that include expert manual stage scoring based on the Rechtschaffen & Kales (R&K) scoring manual. The obtained signals go initially through a pre-processing procedure where sleep stages signals are extracted, normalized, and filtered. The resulting sleep signals are trained using a k-fold cross-validation scheme of 10-folds. Prior to the training and classification process, the LSTM network architecture is built using Uni- and Bi-directional structures to utilize both the forward and backward chains of data sequences. At the end, the developed algorithm performance is evaluated and a complete performance summary table is provided relative to other State-of-the-Art deep learning studies. The performance of this study is evaluated initially without the merging of S3 and S4 sleep stages following the R&K manual, which is considered challenging due to the minor differences between the signals. Then, the performance is evaluated following the recent American Academy of Sleep Medicine (AASM) scoring manual with the merging of the two stages as N3. The developed algorithm achieved higher results using the Bi-directional LSTM. In addition, it achieved the highest accuracy among all other studies in the field with 97.28%. Furthermore, Cohen's kappa and F1-score were more than 72% on average between all sleep stages. According to the confusion matrix, the algorithm successfully classified sleep signals with an overall True Positives percentage of 91.92%. The performance of the algorithm improved following the AASM manual, where the Cohen's kappa value increased from 72.55% to 77.73%. The developed algorithm showed potential in automatic sleep stage classification. Future works include further enhancements on the LSTM algorithm to achieve higher levels of performance.