Classification of sleep disorders and cyclic alternating pattern events using deep learning

Sleep is a fundamental biological process that plays a critical role in maintaining human health, ensuring physiological restoration, and sustaining cognitive functions. Conditions that disrupt the structure, quality, or continuity of this vital process are defined as sleep disorders. Sleep disorders adversely affect both the physical and psychological well-being of individuals and may lead to serious long-term health problems. Moreover, the Cyclic Alternating Pattern (CAP), a microstructural indicator of sleep, provides valuable insights for assessing sleep quality and particularly for detecting sleep disorders. Today, the diagnosis of sleep disorders and the detection of CAP events are primarily based on the manual analysis of physiological signals obtained from polysomnography recordings. However, this process is labor-intensive, time-consuming, and prone to subjective errors. Therefore, developing automated approaches capable of accurately, rapidly, and objectively analyzing sleep disorders and CAP events holds great importance for clinical diagnostic processes. In recent years, artificial intelligence, and especially deep learning (DL) methods, have achieved remarkable success in sleep medicine as in many other complex domains. Numerous studies have been conducted on the automatic detection of sleep disorders and CAP events. Nevertheless, most existing approaches for sleep disorder classification have focused on a single disorder type, have been trained on limited datasets, and have yielded results with low generalizability. Similarly, studies addressing CAP event detection have not sufficiently examined issues such as subtype classification, class imbalance, and explainability. In this thesis study, to address these gaps, DL-based models with high accuracy were developed for multi-class classification of sleep disorders and automatic detection of CAP events. Accordingly, a novel multi-input model named MAF-SleepNet was proposed for sleep disorder classification, while a multi-channel input model named MuRAt-CAP-Net was designed for detecting CAP phase A events and their subtypes. Experimental results demonstrated that MAF-SleepNet achieved accuracy rates of 88.41% and 99.72% for subject-independent and subject-dependent validation strategies, respectively, outperforming existing approaches. MuRAt-CAP-Net, on the other hand, achieved 81.26% and 83.68% accuracy for A-phase detection on balanced and imbalanced datasets, respectively, and 83.34% and 87.31% accuracy for subtype classification, surpassing previous studies.