A Signal Decomposition Model-Based Bayesian Framework for ECG Components Separation

The paper introduces an improved signal decomposition model-based Bayesian framework (EKS6). While it can be employed for multiple purposes, like denoising and features extraction, it is particularly suited for extracting electrocardiogram (ECG) wave-forms from ECG recordings. In this framework, the ECG is represented as the sum of several components, each describing a specific wave (i.e., P, Q, R, S, and T), with a corresponding term in the dynamical model. Characteristic Waveforms (CWs) of the ECG components are taken as hidden state variables, distinctly estimated using a Kalman smoother from sample to sample. Then, CWs can be analyzed separately, accordingly to a specific application. The new dynamical model no longer depends on the amplitude of the Gaussian kernels, so it is capable of separating ECG components even if sudden changes in the CWs appear (e.g., an ectopic beat). Results, obtained on synthetic signals with different levels of noise, showed that the proposed method is indeed more effective in separating the ECG components when compared with another framework recently introduced with the same aims (EKS4). The proposed approach can be used for many applications. In this paper, we verified it for T/QRS ratio calculation. For this purpose, we applied it to 288 signals from the PhysioNet PTB Diagnostic ECG Database. The values of RMSE obtained show that the T/QRS ratio computed on the components extracted from the ECG, corrupted by broadband noise, is closer to the original T/QRS ratio values (RMSE=0.025 for EKS6 and 0.17 for EKS4)