Electrocardiogram-based automatic sleep staging in sleep disordered breathing

A system for electrocardiogram (ECG) based sleep staging in subjects with sleep-disordered breathing is described. Three sleep states are defined: wakefulness(W), REM sleep(R) and non-REM sleep. Features investigated include RR interval, RR standard deviation, RR spectra, respiratory frequency, RR interval differences, and an ECG-derived respiratory signal. A subject specific quadratic discriminant classifier was trained and tested, and yielded an estimated classification accuracy of 71% (Cohen´s κ value of 0.37). When a similar subject-dependent classifier was trained and tested, the estimated classification accuracy dropped to 61% (κ=0.12). For comparison, an electroencephalogram (EEG) based classifier yielded a subject-specific accuracy of 76% (κ=0.51), and subject-independent accuracy of 75% (κ=0.43), indicating that EEG features are robust across subjects. We conclude that the ECG signal provides moderate sleep-staging accuracy, but features exhibit significant subject dependence.