Multi-channel intracardiac electrocardiograms of atrial fibrillation (AF) patients are acquired at the electrophysiology laboratory in order to guide radiofrequency (RF) ablation surgery. Unfortunately, the success rate of RF ablation is still moderate, since the mechanisms underlying AF initiation and maintenance are still not precisely known. In this paper, we use an advanced machine learning model, the Gaussian process latent force model (GP-LFM), to infer the relationship between the observed signals and the unknown (latent or exogenous) sources causing them. The resulting GP-LFM provides valuable information about signal generation and propagation inside the heart, and can then be used to perform causal analysis. Results on realistic synthetic signals, generated using the FitzHugh-Nagumo model, are used to showcase the potential of the proposed approach.
IntroductionThe term atrial fibrillation (AF) is commonly used by cardiologists to denote a family of cardiac arrhythmias characterized by a rapid and unsynchronized contraction of the atria. In spite of its epidemic nature, and the large number of studies performed over the last decades, the mechanisms underlying AF initiation and maintenance are still not precisely known [1][2][3]. One of the leading hypotheses (rotor theory) states that specific areas of the myocardium may be responsible for AF initiation and maintenance [4]. Radiofrequency (RF) catheter ablation intends to terminate AF (and prevent its recurrence) by targeting these arrhythmogenetic areas. However, no consensus has been attained yet on which areas should be ablated, the success rate of a single procedure is still unsatisfactory, and its relative effectiveness w.r.t. the use of antiarrhythmic drugs remains controversial [5][6][7][8].In this paper, we use an advanced machine learning model, the Gaussian process latent force model (GP-LFM) [9][10][11], to describe the intracardiac electrocardiograms (a.k.a. electrograms (EGMs)) acquired during RF ablation at the electrophisiology laboratory. The GP-LFM assumes that the observed EGMs have been generated by some unknown signals, latent forces (LFs), and tries to learn both those LFs, which are modelled as Gaussian processes (GPs) [12], and the input-output mechanism, which is cast within the linear convolution framework. The resulting GP-LFM provides valuable information about signal generation and propagation inside the heart, and can then be used to perform causal analysis between the observed EGMs and the inferred LFs. Note that Granger causality (G-causality) maps were already built in [13,14], whereas a hierarchical G-causality approach was developed in [15,16]. However, none of these approaches takes into account that the observed EGMs have been generated by one or more exogenous sources (i.e., the LFs), as we do here. Results on realistic synthetic signals, generated using the FitzHugh-Nagumo model, are used to showcase the potential of the proposed approach.
Gaussian Process Latent Force ModelLet us assume that we have a mult...