Improving Sleep Apnea Screening with Variational Mode Decomposition and Deep Learning Techniques

Obstructive sleep apnea (OSA), a sleep condition, is characterized by recurrent bouts of irregular breathing. This study uses deep learning (DL) and variational mode decomposition (VMD) techniques to develop a classification system for sleep apnea. VMD is an adaptive signal decomposition technique to simplify complicated signals into a finite number of decomposed intrinsic mode functions (IMF). The baseline systems consist of support vector machine (SVM)-classifiers constructed using statistical features. Each of the ECG recordings is segmented; into one-minute segments. VMD is then performed on each of the one-minute long ECG segments. Statistical features derived from the derived IMFs, time, and frequency domain (TD and FD) segments are used to train SVM classifiers. We then developed convolutional neural networks (CNNs) and evaluated the performance. The CNN models trained with first, second, and third IMFs were the best performing systems. Subsequently, we added the first, second, and third IMFs to reconstruct a denoised ECG signal. CNN model trained with this regenerated ECG signal showed an accuracy, sensitivity, and specificity of 88.187%, 93.128%, and 80.339%, respectively. Subsequently, we extracted bottleneck features (BNF) from the bottleneck layer of the CNN. We trained a smaller dense neural network using the BNFs. When compared to the CNN model, the dense neural network trained with BNF extracted from the regenerated ECG signal (resulting from summing first, second, and third IMFs) gave the best performance with 4.24% 6.84%, and 2.55% improvements in accuracy, specificity, and sensitivity, respectively; with an area under the ROC curve 0.914.