Machine Learning Application in Sleep Disorder Analysis

Abstract

Sleep is a natural state of reduced consciousness and physical activity that is crucial for the body’s circadian rhythm and various physiological processes. Sleep disordered breathing events, such as sleep apnea, can cause cortical arousal during sleep, leading to sleep cycle breaking and sleep fragmentation. Inadequate sleep can jeopardize the immune system and pose a significant risk to health and life. Sleep disordered breathing (SDB) screening and understanding the coupling between cortical arousal and health burden are critical for sleep and health monitoring. Machine learning can be compatible of recognizing patterns in complex data and prediction on future unseen data, which not only can reduce the intensive labor for manual processing but also uncover the new knowledge that people may have not been awareof. In this dissertation, we developed a six-item questionnaire for obstructive sleep apnea (OSA) screening tool using a large general population database - Sleep Heart Health Study(SHHS). Two independent logistic regressions underlie the algorithm of OSA binary prediction by considering two phenotype groups. We evaluated the tool on the SHHS test set (n = 1237) and an independent set Wisconsin Sleep Cohort (WSC) (n = 1120). The model outperformed these questionnaires on both test sets regarding the area under the receiver operating characteristic (AUROC) and the area under the precision-recall curve (AUPRC). The model achieved AUROC (SHHS 1: 0.78, WSC: 0.76) and AUPRC (SHHS 1: 0.72, WSC: 0.74), respectively. A multi-task deep learning algorithm was investigated based on previous study for cortical arousal detection and sleep staging using the single-lead Electrocardiogram (ECG). The model was developed on Multi-Ethnic Study of Atherosclerosis (MESA) dataset of which 1069 full night PSG were included in this study. We analyzed the fairness of the model by comparing the performance between subgroups with different demographics factors including gender, age and ethnicity. The model achieved an AUROC of 0.947 and AUPRC of 0.61 regarding the arousal detection, and Conhen’s κ of 0.68 and overall accuracy of 0.79 for four stage sleep prediction-light(N1/N2), deep(N3/N4), REM and wake. Though numerical differences regarding the aforementioned metrics were observed between different subgroups, the differences were not statistical significant. We also analyzed the intermediate channel outputs that shared by both tasks to explore the features that model have learned from raw ECGs. The results showed that intermediate channel outputs have a strong correlation with instantaneous heart rate on at least 80% of the subjects in testing set. More complicated HRV features were also investigated, though less portion of subjects showed strong correlations. Lastly, we studied the instantaneous association between cortical arousal onsets and heart rate variability (HRV) with the help of deep learning model among the general population. We compared the HRV changes pre-, intra- and post- arousal occurrence using a 25-second window. We also examined the cardiac response difference between different genders (male and female) and different sleep stages(REM or NREM). Significant variations were observed in heart rate and HRVs due to arousal onsets. Most importantly, female showed a more intensive cardiac response to the arousal onsets compared to male subjects which can potentially result in heavier heart burdens and long-term cardiovascular morbidity. More intensive variations caused by arousal were observed in REM, implying the instantaneous elevation of the sympathetic tome which may stress the cardiac function and cause sudden cardiac death.