394 A convolutional neural networks model for the detection of cortical arousals from heart rate

Abstract

Abstract Introduction Cortical arousals are transient events of disturbed sleep that occur frequently in sleep disordered breathing (SDB) and can be used as an indicator of sleep quality. While cortical arousals are typically scored from the electroencephalogram (EEG), arousals are associated with increased sympathetic activity and could therefore be detected from measures of sympathetic activity such as heart rate. Most home sleep test and consumer wearable devices enable continuous recording of heart rate via the electrocardiogram (ECG) or optical heart rate sensors without the inconvenience of EEG electrodes. In this preliminary study, we developed a deep learning-based convolutional neural networks (CNN) model to detect arousals from heart rate. Methods This study included 1,083 polysomnograms (PSGs) from five independent studies (Tucson Children’s Assessment of Sleep Apnea, Mechanisms of Pharyngeal Collapse in Sleep Apnea, Impact of the Arousal Threshold in Obstructive Sleep Apnea, Predicting Successful Sleep Apnea Treatment with Acetazolamide in Heart Failure Patients, Combination Therapy for the Treatment of Obstructive Sleep Apnea) that were scored for arousals according to American Academy of Sleep Medicine scoring rules. These studies included PSGs from both children and adults (ages 6 and above), with most data coming from participants with evidence or diagnosis of SDB. We used the Pan-Tomkins algorithm to detect R-peaks from the raw ECG signal, transformed the peaks into normalized instantaneous heart rate at 1 Hz frequency, and produced arousal probability in 1-second resolution using a simple CNN model. Due to slight asynchrony between the appearance of arousals in the EEG versus the heart rate, all overlaps between model-predicted arousals and manually scored arousals were considered true-positives. Results We evaluated the model on a validation set (n=216). The model achieved a gross area under precision-recall curve score of 0.67 and a gross area under receiver operating characteristics curve of 0.91 Correlation between the number of model-detected and manually scored arousal events was r=0.76. Conclusion This preliminary study demonstrates that a deep learning approach has the potential to accurately detect arousals in home sleep tests and consumer wearable devices that measure heart rate. Support (if any) The study was supported by grant #207-SR-19 from the American Academy of Sleep Medicine Foundation.