Parametric modelling of ECG signal.

Abstract

LIKE MANY other natural signals, the electrocardiograph (ECG) is also a non-stationary signal (GrU~N1ER, 1983). The burst-like QRS feature contributes localised high-frequency components in the ECG signal, making it distinctly non-stationary (WALDO and CHITRAPA, 1991). Although this feature of the QRS wave has helped detection of the wave by filtering/linear prediction (FRIESEN et al., 1990), it makes the modelling of the signal very difficult. Most of the work in modelling a ECG is non-parametric in nature (GRAHAM, 1976; WOMBLE et al., 1977; JALALEDDINE et al., 1990). An attempt to represent a segment of the ECG by the impulse response of a pole-zero model was unsuccessful because of its prohibitively large order (MURTHY et al., 1979). Later, modelling a small segment (about a period) of the ECG by damped sinusoids was found to be superior to the earlier attempt. The method, however, fails to exploit the global nature (e.g. pseudo-periodicity) of the ECG signal (NIRANJAN and MURTHY, 1993). The time-dependent autoregressive (AR)/autoregressive moving average (ARMA) model is the representative of the general class of non-stationary signals (GRENIER, 1983). As such, the model can also be used for the ECG signal. However, the ECG has some distinctive features; its pseudo-periodicity, and different features of the constituent signals (P, QRS and T) representing actions of various parts of the heart (GUYTON, 1985) etc. It would be useful to know how the general timedependent AR/ARMA model is restricted by the special features of the ECG-type signals. We show that the amplitude modulated (AM) sinusoidal signal model, which is a special case of the time-dependent AR/ARMA model, can have the periodicity property, and the model can exhibit a burst-like feature very well, when the modulating signal is an exponential function. We propose that one or more AM sinusoidal signal(s) can be employed to model separately each feature of the ECG signal. The suitability of the developed model is then investigated for the ECG signal using an analysis-by-synthesis technique.