Covariance Normalization and Bottleneck Features for Improving the Performance of Sleep Apnea Screening System

Abstract

Obstructive sleep apnea is a type of sleep disordered breathing (SDB), marked by pauses in breath during sleep. Sleep apnea monitoring devices are extremely expensive and unavailable in rural areas. The focus of this work is to develop a cost-effective sleep apnea screening system based on single channel electrocardiography (ECG) signal. Initially we built a baseline system that used heart rate variability information as input to a CNN classifier. The baseline system performance was evaluated for time domain (TD), frequency domain (FD), and TD and FD HRV features. The baseline model had an overall accuracy of 78.39%, specificity of 70.58% and sensitivity of 86.2%. In an attempt to increase the system performance, two methods were employed. Initially covariance normalization (CVN) was applied to the input features. CVN reduces the noisy factors induced to the input features due to patient specific variations. Subsequently we used neural networks to extract the bottleneck features (BNF) from bottleneck layer of the CNN model. This layer compresses the neural network, allowing the extraction of lower-dimensional information from the network. System performance was evaluated with the BNF extracted from the baseline model with HRV features as input, and also from the baseline model built using normalized HRV features. Upon performance evaluation, it was found that, compared to the baseline model, the BNF extracted from TD and FD HRV features shows a performance improvement of1.39%and BNF extracted from normalized TD and FD HRV features improved the overall accuracy by 1.7%.