Time-frequency modelling and discrimination of noise in the electrocardiogram.

In widely spread home care applications of ECG recorders, the traditional approach to the problem of noise immunity is no longer sufficient. This paper presents a new ECG-dedicated noise removal technique based on a time-frequency noise model computed in a quasi-continuous way. Our algorithm makes use of the local bandwidth variability of cardiac electrical representation and splits the discrete time sequence into two sub-planes. The background activities of any origin (muscle, power line interference, etc) are measured in the regions of the time-frequency plane, situated above the local bandwidth of the signal. The noise estimate on each particular scale is non-uniformly sampled and needs to be extrapolated to the regions where the components of cardiac representation are normally expected. On the lower scales, the noise contribution is computed with the use of square polynomial extrapolation. The time-frequency representation of noise, partially measured and partially calculated, is arithmetically subtracted from the noisy signal, and the inverse time-frequency transform yields a noise-free cardiac representation. The algorithm was tested with the use of CSE database records with the addition of MIT-BIH database noise patterns. The static and dynamic performance of the algorithm is sufficient to ameliorate the signal-to-noise ratio by more than 11 dB.