Comparison of ECG fiducial point extraction methods based on dynamic Bayesian network

Cardiovascular diseases are one of the major causes of mortality in humans. One way to diagnose heart diseases and abnormalities is processing of cardiac signals such as electrocardiogram (ECG) signal. In many ECG analysis, location of peak, onset and offset of ECG waves must be extracted as a preprocessing step. These points are called ECG fiducial points (FPs) and convey clinically useful information. In this paper, we compare some FP extraction methods including three methods proposed recently by our research team. These methods are based on extended Kalman filter (EKF), hidden Markov model (HMM) and switching Kalman filter (SKF). Results are given for ECG signals of QT database. For all proposed methods, the mean of estimation error across all FPs are less than 4 msec (one sample) and their root mean square error are less than 17 msec (almost 4 samples). The proposed methods are also compared with two other methods based on wavelet transform and partially collapsed Gibbs sampler (PCGS). The obtained results by proposed methods outperform two other methods.