Automated LSTM-based sleep stage classification using polysomnographic signal processing techniques

Sleep is a fundamental human physiological activity. It is essential for physical and mental health and adsorbs a significant part of human life. Therefore, the sleep classification into different stages can help to detect disorders related to sleep and it is of great significance for sleep research. Sleep scoring is generally carried out by using polysomnographic (PSG) recordings following the rules of the American Academy of Sleep Medicine. Usually, the analysis of PSG signals is done by experienced doctors. It is carried out by looking at the characteristics of multiple signals (EEG, EOG and EMG) collected from the patient. The objective of this study is to develop a fully automatic method of classification, which replicates the work done by the experts. The approach aims to save doctors time, but above all, it would reduce the intra- and inter-operator variability in stage classification. The ISRUC-Sleep dataset, which contains all-night (around 8 hours) multi-channel PSG signals recorded on patients both healthy and not healthy, was used in this study. The PSG signals were analyzed by different signal processing methods: five signal decomposition algorithms and three signal transforms for the time-frequency analysis. The obtained results were submitted to a feature extraction phase both for time and frequency domains and for the time-frequency domain, to limit the amount of data that was generated by the previous processes. Once the most relevant features, computed by the Minimum Redundancy Maximum Relevance algorithm (mRMR) on 5 s time-step, were selected, a Recurrent Neural Network with Long Short-Term Memory (LSTM) units was applied to classify the sleep stages. The best network classified the five sleep stages (W, N1, N2, N3 and REM) with an overall percentage of correct classification above the average state of the art. Although a misclassification of the N1 class with the N2 class was noted, the results could be considered good. Moreover, the use of EOG and EMG signals revealed necessary for a more accurate classification of sleep phases.