Sleep Stages Classification in a Healthy People Based on Optical Plethysmography and Accelerometer Signals via Wearable Devices

This paper proposes a novel automatic classifier of rapid-eye-movement (REM), non-rapid-eye-movement (NREM, combined light and deep sleep stages (SS)), and wake stages based on wrist photoplethysmography and wrist 3-axis accelerometer data. The quality and robustness of the classifier are achieved by a multilevel learning framework, which consists of features extraction, preprocessing, signal treatment, machine learning, and post-processing modules. Subject-specific Z-score normalization and exponential smoothing are applied to reduce intra-subject and inter-subject variability. The novel features are based on heart rate variability and motion statistics. Leave-one-subject-out cross-validation technique is applied to estimate the quality of the model. Our solution, based on linear discriminant analysis (LDA), shows the Cohen's Kappa coefficient of 0.58 and accuracy of 77 % for the four class (wake, REM, light, deep) classification problem and Cohen's Kappa coefficient of 0.67 and accuracy of 85 % for the three class (wake, REM, NREM) classification problem. Comparison of different ML algorithms, like LDA, support vector machines, and random forest is done. The obtained results of sleep stages classification exceed the analogous state of the art solutions.