SleepWatcher: Detecting sleep apnea/hypopnea syndrome from wearable devices using deep learning

초록

Due to the lack of polysomnography facilities and the cost of testing, it is crucial to detect sleep apnea/hypopnea syndrome (SAHS) using measurable biomedical signals from a wearable device. This paper presents a methodology called SleepWatcher to detect SAHS using biomedical signals from a wearable device. This work addresses the problem of class imbalance and handcrafted feature dependencies of biomedical signals in SAHS detection. Heart rate variability (HRV) and blood oxygen saturation (SpO2) signals are used to train two-dimensional convolutional neural networks, and classify normal, apnea, and hypopnea events. This work experimentally demonstrates that multiple signals can be trained with the same framework without handcrafted features. SleepWatcher consists of two stages. Stage 1 of SleepWatcher classifies normal and abnormal events. The classifier of Stage 1 achieves an accuracy of 89%, and specificity and sensitivity of 89%, and 89% when using only HRV signals, and 86%, 84%, and 89% when using only SpO2 signals. SleepWatcher trains on each of the HRV and SpO2 signals and has improved classification performance when compared to state-ofthe-art methods using each signal. The final results of Stage 1 achieve an accuracy of 87%, and specificity and sensitivity of 83%, and 91%, respectively. Stage 2 classifies apnea and hypopnea events and computes apnea/hypopnea index. The classifier of Stage 2 achieves an accuracy of 97%, and specificity and sensitivity of 99%, and 76%, respectively. With SleepWatcher, SAHS diagnosis could become an out-of-hospital procedure with satisfactory performance in terms of accuracy, specificity, and sensitivity.

키워드