Separating Cardiac and Muscular Ecg Components Using Adaptive Modelling in Time-frequency Domain

The paper presents an adaptive time-frequency denoising algorithm. Other denoising methods use a very general probability-based noise model and as general-purpose algorithms rarely consider the a priori knowledge about the signal. Main novelty of the proposed algorithm is the running quasi-continuous scalo-temporal model of background activity built and subtracted from the ECG in order to yield a rectified representation of cardiac action. Our algorithm is based on the P, QRS and T wave borders automatically detected in the ECG and adapts the general information on expected local signal bandwidth to each particular heartbeat. This leads to determine timefrequency regions containing cardiac representation. The complement is assumed to contain only the background activity representation and thus these values can be picked-up directly to the timescale model of noise. For the remaining part of scalo-temporal surface the noise is interpolated with cubic splines in each scale independently and than extrapolated to lower scales. The numerical tests performed with use of artificially noise-affected test signals reveal highly discriminative properties of the method. The amount of removed noise varies from 65% to 90% (SNR increased by 6.5 dB and 11.6 dB respectively) depending on input noise level. The timefrequency noise model is quasi-continuous and adapts to the physiological changes of muscular activity using maximum available real data points. The use of the standard bandwidth function is arbitrary for the ECG, but allows the user to adapt the method to other signals of variable information density.