A comparison of three ECG-derived respiration methods for sleep apnoea detection

Abstract

This study compared three different methods of estimating ECG-derived respiration (EDR) from full overnight single lead ECG signals. The first EDR method was based on calculating an area enclosed by each QRS complex (EDR-A), the second method was an amplitude demodulation method (EDR-B), and the third method used a novel algorithm to provide fast calculation of principal component analysis (PCA) to estimate the EDR (EDR-P). The ‘Apnea-ECG’ MIT-BIH PhysioNet dataset was used to enable the comparisons. An assessment of the Pearson correlation of each EDR method with selected segments of chest respiratory effort was performed. Following this, a comparative assessment of the performance of each EDR method to automatically detect epochs of sleep apnoea was performed. Identical features were derived from the three EDR signals and processed with automated classifiers developed by the linear-discriminant analysis and the extreme-learning machine algorithms. The pooled average correlation for the EDR-A, EDR-B and EDR-P methods against the chest respiratory effort signal were 0.82, 0.79, 0.73. The classification accuracy for detecting one-minute epochs of sleep apnoea was 84%, 79% and 81% with the LDA classifier. We conclude that the EDR-A method provides best performance of the methods.