Improving the Performance of ECG based Sleep Apnea Screening System using Signal Decomposition and Decision Fusion

Obstructive sleep apnea (OSA) is characterized by recurrent episodes of interrupted or reduced airflow during sleep that is brought on by blockage of the upper airway. OSA in adults needs to be diagnosed and treated to ensure quality sleep and good health. Polysomnography (PSG), which is typically used for diagnosing sleep apnea (SA), is available only in specialty hospitals and is not easily accessible to people in rural areas. The primary aim of this work is to develop an affordable sleep apnea screening system using electrocardiography (ECG) signal as its input, thus extending specialty healthcare to rural communities. In this study, a convolutional neural network (CNN) based apnea screening system is proposed, the baseline system giving an accuracy of 75.04%, specificity of 88.31%, a sensitivity of 54.01%, and area under receiver operative characteristics (AUROC) 0.7116. The CNN was then used for deriving bottleneck features from the ECG signal, which is then given as input to a dense neural network (DNN). Further, we explored two decomposition techniques, variational mode decomposition (VMD) and wavelet decomposition (WD). The intrinsic mode functions (IMFs) obtained from the VMD as well as the various wavelet levels obtained from the WD were evaluated individually for their performance. The best-performing IMFs from VMD are combined to reconstruct the time domain signal to be fed as input to the CNN resulting an improvement in the system performance. Similarly, the best-performing wavelet levels from WD are also combined to obtain another system. Subsequently, the decision scores of these two systems are combined through the decision fusion technique with weights empirically derived for the best performance. The final system gave an accuracy of 79.58%, a sensitivity of 82.31
^%
, and a specificity of 77.86% with an AUROC of 0.8009, an improvement of 4.54
^%
in accuracy and 0.09 increase in AUROC score with respect to the baseline system.