SC3: self-configuring classifier combination for obstructive sleep apnea

Abstract

Obstructive sleep apnea is considered to be one of the most prevalent sleep-related disorders that can affect the general population. However, the gold standard for the diagnosis, polysomnography, is an expensive and complicated process that is commonly unavailable to a large group of the population. Alternatively, automatic approaches have been developed to address this issue. One of the goals of this research is to perform the classification of the apnea events with the lowest possible number of sensors. Therefore, the blood oxygen saturation signal was employed in this work since it is correlated with the occurrence of apnea events and it can be measured from a single noninvasive sensor. The events detection was performed by a combination of classifiers. However, choosing the type of classifier to combine and select the most relevant features for each classifier is considered to be a well-known problem in the field of machine learning. A self-configuring classifier combination technique based on genetic algorithms was developed for multiple classifiers and features selection which was tested along with different databases and input sizes. The best performance for obstructive sleep apnea detection was achieved using maximum voting independent feature selection with 1 min time window having the best sensitivity of 82.48% similar database in the literature. This model was later tested on another database for cross-database accuracy. With an average accuracy of 91.33%, the system proved its capabilities for clinical diagnosis since the model was developed and validated with both subject and database independence.