Detection of obstructive sleep apnea using nonlinear time series analysis of speech signals and intelligent decision making methods

Obstructive sleep apnea (OSA) is a prevalent sleep disorder. Due to the challenges encountered in the application of polysomnography (PSG), which is the gold standard diagnostic method for OSA, research in this field has been focused on the development of various approaches that can be used as alternatives to PSG. Detection of OSA using speech/voice signals recorded while individuals are awake has become a significant area of research in recent years. In previous studies, typically vowels and some nasal consonants were examined with classical voice analysis, but the results did not reach success levels that could be translated into clinical practice. This thesis study was designed with the idea that exploring the chaotic behavior inherent in voice through nonlinear analysis approaches could yield effective results in detecting OSA by capturing dynamics that classical voice analyses cannot adequately reveal. Methods known as nonlinear time series analysis, based on chaos theory, were employed, incorporating various features and machine learning to detect OSA and determine its severity (mild, moderate, severe apnea). In this study, 32 voices with the potential to induce changes in the voices of individuals with OSA were identified, and recorded from 141 subjects. The analysis results were evaluated for vowels, consonants, and all voices, and classification studies were conducted, with the best results obtained for consonants. In the analysis of consonants, 336 features (28 voices × 12 features) were calculated for each subject. Five features were selected for healthy/OSA classification, and 14 features were selected for apnea severity classification, using the ANOVA feature selection method. By employing various configurations of K-nearest neighbors (K-NN) and support vector machines (SVM) classifiers, the study achieved a 95.1% accuracy in detecting healthy/OSA and an 82% accuracy in classifying OSA severity. The studies demonstrated that the presence of OSA and the OSA severity could be determined within approximately 15 minutes, consistent with PSG results (simple snoring, mild, moderate, and severe OSA), using a small number of nonlinear features calculated from various sound samples. As a result, the study achieved the highest healthy/OSA classification accuracy rate in the literature. Additionally, it is the first study in the literature to determine the OSA severity from speech, achieving a very high accuracy.