Intracardiac Activity Characterization in Atrial Fibrillation using Deep Learning

Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia worldwide and is associated with increased risk of stroke, heart failure, and mortality [1]. The prevalence of AF is expected to rise markedly in the coming decades, driven by population aging, the growing incidence of cardiovascular comorbidities, and improved diagnostic capabilities [2]. Consequently, there is an urgent need for more effective early diagnostic and therapeutic strategies that can
improve patient outcomes and support personalized clinical decision-making.
Restoring sinus rhythm is a primary clinical objective, often pursued through catheter ablation [3]. Ablation strategies have traditionally focused on the initiating triggers, most notably the pulmonary veins (PV), which are hypothesized as the dominant source of ectopic activity in paroxysmal AF. However, despite technological advances in ablation techniques, long-term success rates remain suboptimal, especially in patients with persistent or long-standing persistent
AF [4, 5].
The limited efficacy of conventional ablation strategies has motivated extensive research into the electrophysiological mechanisms sustaining AF [6]. Several mechanistic frameworks have been proposed, including the presence of localized drivers such as rotors, focal sources, or regions of functional reentry, as well as the multiple wavelet hypothesis involving complex, self-sustained reentrant activity [4, 5, 7] theoretically measurable from intracardiac electrograms (EGMs).
Aiming to localize the sources of these mechanisms for ablation, current invasive mapping techniques enable the acquisition of intracardiac EGMs with high temporal resolution, providing detailed local information on atrial activation, voltage, and conduction [8]. Nevertheless, these methods suffer from intrinsic limitations, including restricted spatial resolution, incomplete atrial coverage, and dependency on catheter contact, prolonged procedure times, and difficulty
in capturing the full three-dimensional dynamics of the atria [9].
Non-invasive approaches based on electrocardiographic imaging (ECGI) have been developed to address some of these limitations by reconstructing atrial electrical activity from body surface potential measurements (BSPMs) [10]. ECGI offers the potential to provide global maps of atrial activation noninvasively, thereby complementing invasive procedures. However, classical ECGI methods are fundamentally constrained by the ill-posed nature of the inverse problem of
electrocardiography, which requires strong regularization and relies heavily on accurate patientspecific torso–heart geometries obtained from medical imaging [11]. Consequently, conventional ECGI reconstructions often suffer from over-smoothing and loss of fine-grained electrophysiological information, limiting their clinical utility for functional mapping during AF [11–14]. These challenges are amplified in atria, which faces weaker signals, more complex and variable activation,
and fewer established methods than ventricular ECGI. [15]. Overall, these limitations underscore the need for new tools that can globally capture AF dynamics to improve mechanistic understanding, guide invasive mapping, and enable more accurate ablation targeting with better outcomes.
Advances in deep learning have recently introduced new possibilities for non-invasive characterization of cardiac electrophysiology, offering a data-driven alternative to classical ECGI methods that rely on explicit prior anatomical information and mathematical regularization [14, 16, 17]. In particular, deep learning has emerged as a promising framework to address three interrelated challenges that are central to AF research: the detection of sustaining drivers, the non-invasive
estimation of EGMs, and the modeling and synthesis of atrial electrical signals. Unlike traditional model-driven approaches, deep learning methods learn complex, non-linear relationships directly from data, making them well suited to capture the multiscale, non-stationary, and highly heterogeneous dynamics that characterize AF.
Early efforts in rotor detection focused on the mechanical identification of AF triggers through phase mapping, based on both invasive recordings [18] and traditional ECGI techniques [16, 19]. More recently, machine learning and deep learning methods have been proposed for AFdriver detection, circumventing explicit mechanistic metrics and instead leveraging data-driven feature extraction [20–22]. Beyond driver localization, the direct non-invasive estimation of intracardiac EGMs constitutes a more granular and information-rich objective. To this end, a variety of deep learning architectures have been proposed as alternatives to traditional ECGI. Convolutional neural networks (CNNs) have demonstrated strong performance in capturing spatial patterns in torso potential maps [22–24], while recurrent architectures have shown effectiveness in reconstructing local EGMs with preserved temporal and spectral characteristics, particularly in the context of AF [23, 25, 26]. A major bottleneck in the application of deep learning to atrial electrophysiology is the limited availability of large, well-annotated intracardiac datasets, particularly for AF. To mitigate this challenge, deep learning–based generative modeling has emerged as a promising strategy for synthetic ECG generation. Among the most widely explored approaches are variational autoencoders (VAEs) [27, 28].