Deep Feature Extraction and Early Prediction of Obstructive Sleep Apnea Events

Abstract

Obstructive sleep apnea (OSA) is a common sleeping breathing disorder characterized by interruptions in breathing or obstructions in the airway. An early prediction of OSA may help in avoiding the disorder’s symptoms by identifying events before they happen. In this regard, we proposed a methodology for OSA prediction using both convolutional neural networks and traditional machine learning approaches. For this purpose, 30-second pre-apnea and non-apnea segments of electrocardiogram (ECG) recordings were extracted for various leading times and 2D scalogram images representing the time-frequency characteristics were generated from each segment. Deep features extracted from scalogram images using a modified residual network were fed separately into a support vector machine and two ensemble classifiers, namely random subspace k-nearest neighbors (kNN) and random subspace discriminant classifiers. The subspace kNN classifier outperformed other classifiers and achieved performance results up to 86,97% accuracy, 88,19% sensitivity, 84,26% specificity, and 85,70% positive predictive value. These results suggest that using machine learning approaches to classify deep features of single lead ECG scalogram images may improve prediction performance. Ultimately, the proposed method can be used as an useful approximation to identify impending OSA events