Artificial Intelligence for Assessing the Correlation Between Sleep Apnoea and Comorbidities

Abstract

Sleep apnoea (SA) is a chronic sleep related breathing disorder characterized by partial or complete airway obstruction, known as Obstructive Sleep Apnoea (OSA), or when the brain does not communicate the proper signals to the muscles responsible for breathing, known as Central Sleep Apnoea (CSA) in sleep leading to pauses in breathing. The consequences of SA in patients range from conditions such as excessive daytime sleepiness, insomnia and poor mental health to severe co-morbidities such as coronary heart disease, strokes, and diabetes. The clinically recommended treatments for mitigating OSA are continuous positive airway pressure, weight management and stress reduction. The gold standard method of diagnosis is the polysomnography test (PSG) performed in a sleep laboratory and estimates the severity of the condition by assessing the respective oximetry, snoring, body movements, oronasal airflow, electrocardiogram, electroencephalogram, electro-oculogram and electromyogram measurements. Interestingly, these physiological measures reveal additional information regarding the progression of other relevant conditions such as mental health disorders and cardiovascular complications. The expensive, time consuming and labour-intensive nature of PSG, and the lack of regular monitoring in patients' daily lives with existing solutions motivates the development of clinical support for enhanced prognosis. The goal of this thesis is to provide a comprehensive study for classification of two types of SA, OSA and CSA, by leveraging predictor variables easily acquired by wearable devices or electronic health records. More specifically, we assess the interplay between SA and multiple potential influencers such as depression, anxiety, cardiac activity, sleep habits, sleep physiology, physical measurements with machine learning, deep learning, and statistical methods. The performance measures of the various constructed algorithms for classifying SA are reported using standard evaluation metrics. Furthermore, post-hoc feature importance analysis is used for model decision interpretation with Shapley values, and calibration is conducted to align model prediction probability with relative frequency of the disease in the dataset. Finally, we implement a model on an FPGA for sleep staging in a sleep-disordered population.