Introduction: Arousals during sleep give deep insights into the pathophysiology of sleep disorders and sleep quality. Medical gold standard for detecting arousals is a time-consuming process manually performed by a trained expert. Arousals are detected visually in the electroencephalogram and electromyogram, which are part of the polysomnography. Measuring these biosignals requires a stationary setup and is uncomfortable for the patient. As arousals relate to the autonomic nervous system, they also reflect in the electrocardiogram and heart rate in general, which is therefore a promising alternative biosignal for arousal detection. In this study, we developed a deep learning model for automatic detection of sleep arousals from heart rate.
Method: We developed our algorithm using electrocardiograms and arousal labels from the Sleep Heart Health Study. Firstly, we derived RR intervals from the electrocardiograms and interpolated them into a 4 Hz signal. Excluding low quality signals and records without arousal labels, this resulted in a database of full-night measurements from 5323 subjects. 1003 of them were held-out as test data. Next, we developed a convolutional neural network (CNN) for end-to-end event detection. Model output was a continuous arousal probability with a frequency of 1 Hz. The final architecture was optimized using an extensive grid search and cross validation.
Results: The optimization resulted in a twelve-layer CNN that achieved a Cohen's kappa of 0.47, an area under the precision-recall curve of 0.53, and an area under the receiver operating characteristic of 0.86 on hold-out test data.
Conclusion: This study demonstrates the ability of machine learning to detect arousals during sleep from heart rate. As our approach uses only the heart rate, it is potentially transferable to other signals, e.g. the photoplethysmogram.