An adaptive particle weighting strategy for ECG denoising using marginalized particle extended Kalman filter: An evaluation in arrhythmia contexts

Model-based Bayesian frameworks have a common problem in processing electrocardiogram (ECG) signals with sudden morphological changes. This situation often happens in the case of arrhythmias where ECGs do not obey the predefined state models. To solve this problem, in this paper, a model-based Bayesian denoising framework is proposed using marginalized particle-extended Kalman filter (MP-EKF), variational mode decomposition, and a novel fuzzy-based adaptive particle weighting strategy. This strategy helps MP-EKF to perform well even when the morphology of signal does not comply with the predefined dynamic model. In addition, this strategy adapts MP-EKF's behavior to the acquired measurements in different input signal to noise ratios (SNRs). At low input SNRs, this strategy decreases the particles' trust level to the measurements while increasing their trust level to a synthetic ECG constructed with the feature parameters of ECG dynamic model. At high input SNRs, the particles' trust level to the measurements is increased and the trust level to synthetic ECG is decreased. The proposed method was evaluated on MIT-BIH normal sinus rhythm database and compared with EKF/EKS frameworks and previously proposed MP-EKF. It was also evaluated on ECG segments extracted from MIT-BIH arrhythmia database, which contained ventricular and atrial arrhythmia. The results showed that the proposed algorithm had a noticeable superiority over benchmark methods from both SNR improvement and multiscale entropy based weighted distortion (MSEWPRD) viewpoints at low input SNRs.