Determination of obstructive sleep apnea with nonlinear analysis of speaking signals

Obstructive Sleep Apnea (OSA) is one of the most common sleep disorders. Polysomnography is a gold standard diagnostic method for OSA. OSA can be diagnosed as a result of recording many physiological parameters of patients during a night's sleep and scoring these records. Due to the difficulties encountered during its implementation and the high cost, the studies for the development of simplified approaches instead of Polysomnography in OSA detection have gained importance in recent years. Studies involving the evaluation of physiological parameters and snoring sounds, which are one of the OSA markers, that patients can record on their own at home, have shown some success in the detection of apnea. However, in these methods, the disadvantage of recording night long could not be eliminated. In recent years, studies, that take into account the commonness of the respiratory and vocal tract, focus on OSA detection with the approach of evaluating the speech sounds containing a few minutes of vowels recorded while patients are awake by using the classical sound analysis. In this study, unlike the studies in the literature, it was aimed to detect OSA with high accuracy using consonant sounds produced by the muscles (art palate, velum, uvula or tongue root) that shake during sleep and block the airway in OSA patients. In the study, the characteristics revealing the nonlinear nature of consonant {/ca/, /ci/, /cı/, /cu/, /ga/, /gi/, /gı/, /gu/, /ha/, /hi/, /hı/, /hu/, /ka/, /ki/, /kı/, /ku/, /na/, /ni/, /nı/, /nu/, /sa/, /si/, /sı/, /su/} and vowel {/a/, /i/, /ı/, /u/} sounds obtained from 20 OSA patients and 20 healthy subjects were calculated, and their OSA detection performances were evaluated with the K Nearest Neighbors (K-NN) and Support Vector Machines (SVM) classifiers. Firstly, classification is made only for vowels, only consonants and all sounds with nonlinear features. Then, some properties used in classical sound analysis were added to the nonlinear features and the classification procedures were repeated. OSA detection accuracies obtained with nonlinear features were found as 97.5%, 97.5% and 100% (with K-NN); 95%, 97.5% and 96.25% (with DVM), for vowels, consonants and all sounds, respectively. In classifications where both non-linear and classical sound analysis features are used together, K-NN classification accuracy increased only in vowel sounds and reached 100%, while other groups almost did not change. DVM accuracies in this application were slightly increased for vowels and all voices (98.75%), but did not change for consonant voices (97.5%). These results indicate that nonlinear features are distinctive for consonant sounds. The obtained results show that nonlinear features provide much higher OSA detection performance for both consonant and vowel sounds than those reported in the literature with classical sound analysis. As a result, in this thesis study, we reached higher values with a few minutes of voice recording to be taken from OSA patients and nonlinear time series analysis than those achieved by previous studies, using traditional sound analysis. It is anticipated that the system proposed in the study can successfully make pre-diagnosis of OSA.