Steps toward subject-specific classification in ECG-based detection of sleep apnea.

This study deals with ECG-based recognition of sleep apnea in epochs of 1 min duration using spectral- and correlation-based features extracted from the modulation of QRS amplitude, respiratory myogram interference and RR intervals. On a database comprising 140 simultaneous recordings of polysomnograms (PSGs) and 8-lead Holter-ECGs, it is shown that a single-parameter ROC threshold classification can achieve high detection rates up to 81.0% sensitivity and 85.6% specificity. Still, individual accuracy may be low, and the improvement employing feature combination by means of second order polynomial classifiers is only marginal. We speculate that individual differences, like co-morbidities, and even intra-individual confounding factors, like nocturnal changes in body position (BP), are major reasons for the difficulties to significantly raise the detection rate using multivariate techniques, which is evident in virtually all papers on that subject. Using the BP information in the PSG, we show a potential benefit for individualized single-feature classifiers by comparing the maximally achievable individual and global accuracy when either one optimal global threshold for the total dataset, individual threshold values for each subject or individual thresholds for each BP are applied. We developed an ECG-based BP segmentation algorithm and finally suggest a potential strategy to derive individually optimized subject-specific threshold values.