Assessment of different methodologies to include temporal information in classifying episodes of sleep apnea based on single-lead electrocardiogram

Automated analysis of sleep apnea based on single-lead electrocardiogram would make screening and diagnosis much more accessible. Over the years, several algorithms have been proposed in the literature. In most of them, one or several temporal averaging techniques are used to improve classifier performance. A comprehensive comparison between those techniques however has never been published. Four different temporal averaging techniques, as well as overlapping of segments, were independently assessed using a database of 70 night-time recordings, originally released for the Computers in Cardiology challenge in 2000. Classification was performed with an LDA classifier. Multiple problem-specific feature sets of 10 features were selected out of a complete set of 304 using a two-step approach. Averaging classifier input features over neighboring segments led to the highest agreement values on the test set, outperforming the best automatic entry during the original competition (90.4% vs 89.4%). When combining classifier output values, an odd amount of segments should be used. Calculating features on larger segments (> i-min) led to the worst results, possibly explained by its higher susceptibility to noise. Overlapping of segments improved overall agreement by about 1%.