A graph signal processing framework for atrial activity extraction

Atrial fibrillation (AF) is a common cardiac arrhythmia and its mechanisms are not yet fully understood. Analyzing atrial epicardial electrograms (EGMs) is important to understand the mechanisms underlying AF. However, when measuring the atrial activity (AA), the electrogram is commonly distorted by the far-field ventricular activity (VA). During sinus rhythm, the AA and the VA are separated in time. However, the VA often overlaps with the AA in both time and frequency domain during AF, complicating proper analysis of the AA. Unlike traditional methods, this work explores graph signal processing (GSP) tools for AA extraction in EGMs. Since EGMs are time-varying and non-stationary, we put forward the joint graph and short-time Fourier transform to analyze the graph signal along both time and vertices. It is found that the temporal frequency components of the AA and the VA exhibit different levels of spatial variation over the graph in the joint domain. Subsequently, we exploit these findings to propose a novel algorithm for extracting the AA based on graph smoothness. Experimental results on synthetic and real data show that the smoothness analysis of the EGMs over the atrial area enables us to better extract the AA.