Global Waveform Delineation For RR Series Estimation: Detecting The Sleep Apnea Pattern
C. Marchesi, M. Paoletti, S. Di Gaetano
University of Firenze
Sleep Apnea (SA) forces on the time series of RR intervals (tachogram) a behavior very close to the time domain pattern associated with the Valsalva maneuver apnea. Thus, in principle, methods of analysis of the tachogram obtained during autonomic tests can be extended to documenting SA episodes. In the experience of the authors a contour plot in the time-frequency domain is one of the examples of methods possibly well suited to assist in visual SA detection. This paper is rather dealing with automatic quantitative detection of SA episodes. According to the above SA interpretation, we propose an algorithm for SA time domain extraction from a single lead ECG based on following scheme: a) Generating a tachogram as accurate and precise as possible, even when it is derived from a low sampling rate sampled signal; b) Improving signal to noise ratio, SA pattern playing as signal and regular rhythm as noise; c) Detecting SA episodes on the preprocessed tachogram. As far as point a) we propose a model based beat delineation approach. The QRS global curve fitting provided by the two parameters Gamma density function is a key-solution to obtain an accurate beat delineation, even in poorly sampled signals. In fact, the mathematical model fitting the raw data improves time reference stability since it captures and learns the shape of the wanted event; b) Tachogram preprocessing is basically done with a double moving average process: the first provides to de-trend data, giving the noisy tachogram an amplitude reference, the second filters the high frequency regular rhythm component; c) SA detection performance results to be improved by a final transformation aimed at enhancing discrimination between regular rhythm and SA: from the curve-length concept a non linear estimation of the SA pattern provides the criterion for a threshold based detection. The analysis of the ECG records of the CinC challenge obtained (for the time being) the score of 27/30, ref#20000428.111921 (.Dg99StC).