A New Approach for Estimating Sleep Efficiency from Electromyography Signal

The diagnosis of many sleep disorders mostly depends on the accurate estimation of sleep efficiency (SE) that results from the calculation of total sleep time (TST). By using electromyography signal and machine learning techniques, this study aims to develop a method for estimating SE by accurately predicting TST and total wake time (TWT). A dataset of 97 individuals suspected to have sleep issues was employed, with 81 cases serving as training data and 16 as testing data. After pre-processing, nine different features with 511 unique combinations were utilized to find out the best indicator of sleep duration. A total of 12 machine learning algorithms were applied to evaluate performance metrics such as sleep time accuracy, sensitivity, specificity, and F1 score. In addition, TST, TWT, and hence SE were automatically calculated from the classified sleep/wake state. The overall optimal result was obtained by the model ‘Decision Tree with a depth of 3’ for the feature combination of the ‘fast Fourier transform Magnitude Sum’ and the ‘Mean’. Epoch-by-epoch analysis resulted in sleep time accuracy, sensitivity, specificity, and F1 score of 91.16%, 89.09%, 62.19%, and 85.32%, respectively. In addition, this new approach resulted in an overall relative detection rate of 98.7% for estimating SE. This approach accomplished better results than the existing studies and, hence, would be applicable in diagnosing sleep-related problems, especially sleep apnea.