Photoplethysmographic-based automated sleep-wake classification using a support vector machine

Sleep quality has a significant impact on human mental and physical health. The detection of sleep-wake states is thus of paramount importance in the study of sleep. The gold standard method for sleep-wake classification is multi-sensor-based polysomnography (PSG) which is normally recorded in a clinical setting. The main drawbacks of PSG are the inconvenience to the subjects, the impact of discomfort on normal sleep cycles, and its requirement for experts' interpretation. In contrast, we aim to design an automated approach for sleep-wake classification using a wearable fingertip photoplethysmographic (PPG) signal.

Time domain features are extracted from PPG and PPG-based surrogate cardiac signals for sleep-wake classification. A minimal-redundancy-maximal-relevance feature selection algorithm is employed to reduce irrelevant and redundant features.

A support vector machine (SVM)-based supervised machine-learning classifier is then used to classify sleep and wake states. The model is trained using 70% of the events (6575 sleep-wake events) from the dataset, and the remaining 30% of events (2818 sleep-wake events) are used for evaluating the performance of the model. Furthermore, the proposed model demonstrates a comparable performance (accuracy 81.10%, sensitivity 81.06%, specificity 82.50%, precision 99.37%, and F score 81.74%) with respect to the existing uni-modal and multi-modal methods for sleep-wake classification.

This result advocates the potential of wearable PPG-based sleep-wake classification. A wearable PPG-based system would help in continuous, non-invasive monitoring of sleep quality.