Estimation of the apnea-hypopnea index from epoch-based classification of the ECG using modulations of QRS area and respiratory myogram interference

Most approaches for ECG-based detection of sleep apnea classify fixed-duration epochs (here: 60s) in a binary way with respect to the presence or absence of respiratory events. Identifying time-locked modulations in QRS-amplitude and respiratory myogram interference with a single feature, this paper presents a new method to transform such results into an estimate of the apnea-hypopnea-index (AHI). The basic idea is to translate the local period duration of the event-related quasi-periodic oscillations into a weighting factor for each epoch's result. We present a binary and a ternary strategy (including a borderline class) for identification of patients with an AHI ≥ 15. Results on a large (N = 140) and representative clinical sample indicate 100% sensitivity and 86.4% accuracy for the binary strategy. The ternary strategy achieves 95.1% accuracy for a subset of 83.6% of the recordings. The remaining 16.4% of the sample are classified as `boderline cases'. Additional validation on the independent Physionet Apnea ECG Database (N = 69) resulted in perfect separation of the apnea group from the control cases. We conclude that it is possible to provide a physiologically meaningful estimate of the AHI from epoch-based classifications of the ECG: It promises robust detection of apnea patients in various clinical scenarios such as screening of Holter ECGs.