Automated detection of sleep apnea using sparse residual entropy features with various dictionaries extracted from heart rate and EDR signals

• Published in: Computers in biology and medicine Vol. 108; pp. 20 - 30
• Main Authors: Viswabhargav, ChS.S.S., Tripathy, R.K., Acharya, U. Rajendra
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.05.2019; Elsevier Limited
• Subjects: Accuracy; Annotations; Apnea; Automation; Basis functions; Classifiers; Cluster analysis; Clustering; Decomposition; Dictionaries; EDR and HRV signals dictionaries; EKG; Electrocardiography; Entropy; Feature extraction; Fourier transforms; Heart rate; Principal components analysis; Radial basis function; Respiratory tract; Sensitivity; Sleep; Sleep apnea; Sleep apnea detection; Sleep disorders; Sparse residual entropy; Support vector machines; SVM classifier; Vector quantization; Sleep apnea detection; EDR and HRV signals dictionaries; Sparse residual entropy; SVM classifier
• ISSN: 0010-4825; 1879-0534; 1879-0534
• DOI: 10.1016/j.compbiomed.2019.03.016

Sleep is a prominent physiological activity in our daily life. Sleep apnea is the category of sleep disorder during which the breathing of the person diminishes causing the alternation in the upper airway resistance. The electrocardiogram derived respiration (EDR) and heart rate (RR-time-series) signals are normally used for the detection of sleep apnea as these two signals capture cardio-pulmonary activity information. Hence, the analysis of these two signals provides vital information about sleep apnea. In this paper, we propose the novel sparse residual entropy (SRE) features for the automated detection of sleep apnea using EDR and heart rate signals. The features required for the automated detection of sleep apnea are extracted in three steps: (i) atomic decomposition based residual estimation from both EDR and heart rate signals using orthogonal matching pursuit (OMP) with different dictionaries, (ii) estimation of probabilities from each sparse residual, and (iii) calculation of the entropy features. The proposed SRE features are fed to the combination of fuzzy K-means clustering and support vector machine (SVM) to pick the best performing classifier. The experimental results demonstrate that the proposed SRE features with radial basis function (RBF) kernel-based SVM classifier yielded higher performance with accuracy, sensitivity and specificity values of 78.07%, 78.01%, and 78.13%, respectively with Fourier dictionary and 10-fold cross-validation. For subject-specific or leave-one-out validation case, the SVM classifier has sensitivity and specificity of 85.43% and 92.60%, respectively using SRE features with Fourier dictionary (FD). •A new method based on the analysis of heart rate and EDR signals is proposed for the detection of sleep apnea.•The sparse residual entropy (SRE) features are evaluated from heart rate and EDR signals using various dictionaries.•Six dictionaries such as Fourier dictionary, wavelet dictionary, Eigen dictionary, cosine dictionary, apnea specific learned dictionary and normal specific learned dictionary are used.•For subject-specific validation scenario, the proposed method has sensitivity and specificity of 85.43% and 92.60%, respectively.