International Journal of Medical Discoveries

Background: Sleep specialists often identify sleep stages manually by visually inspecting human electro-biological signals. This manual process is tedious, requires highexperience, and is error-prone. Automation procedures may solely use Electroencephalogram (EEG) signals, which may not be able to capture sleep-related data in other physiological pathways. Objective: This study aims to analyze whether using multimodal physiological measures, EEG, Electrooculography (EOG), and Electromyography (EMG), in conjunction with entropy-based features, can enhance the six-class automatic classification of sleep stages. Method: A subset of 15 polysomnographic records of the Sleep-EDF Expanded database was used to obtain data. Having eliminated epochs where there were no values of the entropy, there were 2918 epochs to analyze. Ten features wereextracted, including Range Entropy (RangeEn) and Sample Entropy (SampEn) of EEG (Fpz-Cz, Pz-Oz), EOG, and EMG. A Linear Discriminant Analysis (LDA) classifier was used and trained with the help of Leave-One-Out Cross-Validation (LOOCV). Sensitivity, specificity, and accuracy were used to measure performance. Results: The multimodal LDA classifier reached an accuracy of 84%, with sensitivity at 64% and specificity at 90% after 2918 epochs. Best performance was in Wake stages (82% sensitivity, 99% specificity) and moderate in REM (61% sensitivity, 88% specificity). Challenges persisted in classifying deep sleep stages S3 and S4 with sensitivities of 55% and 69%, respectively. Significant predictors included entropy parameters, particularly beta-band RangeEn. Excluding EOG and EMG entropy features lowered overall accuracy to 79% and reduced sensitivities for Wake and REM stages, highlighting the importance of multimodal entropy information. Conclusion: These results indicate the usefulness of multimodal PSG signals in automated sleep scoring. Future research should evaluate the generalization to larger cohorts and attempt to evaluate how robust it is to inter-subject variability.