A Preliminary Study on OSA Severity Levels Detection by Evaluating Speech Signals Nonlinearities With Multi-Class Classification

Diagnosis of obstructive sleep apnea (OSA) from speech has become a popular research area in recent years, which can be an alternative way to the application difficulties in polysomnography (PSG). The promising results obtained in our previous study, in which we tried to detect apnea using nonlinear analysis of speech, gave rise to the thought that it is possible to detect OSA and OSA severity by diversifying speech samples and nonlinear features. The principal aim of this study, for the first time in the literature, is to detect the OSA severity levels as mild, moderate, and severe as in the clinic use (multi-class classification) using nonlinear analyses of speech while the patient is awake. In addition, healthy/OSA classification (binary classification) was also carried out. The feature selection method of ANOVA was applied to 336 features (28 voices
$\times12$
features) for each subject, 14 and 5 features were used in multi-class and binary classifications, respectively. As a result of the classifications made with various KNN and SVMs models, the best results were obtained by SVMs in both classifications for OSA severities (with one-vs-all classification scheme and the Gaussian kernel) and OSA detection (with the quadratic kernel) as 82% and 95.1% accuracies, respectively. The proposed study showed that OSA and OSA severity can be determined with the small number of nonlinear features calculated from a few different speech samples, in nearly 15 minutes, consistent with PSG results (simple snorer, mild, moderate, and severe OSA). In conclusion, the highest OSA/healthy classification accuracy rate in the literature was achieved. Furthermore, OSA severity detection in four-class performed quite well as a preliminary study.