Artificial intelligence software standardizes electrogram-based ablation outcome for persistent atrial fibrillation

Abstract

Introduction: Multiple groups have reported on the usefulness of ablating in atrial regions exhibiting abnormal electrograms during atrial fibrillation (AF). Still, previous studies have suggested that ablation outcomes are highly operator- and center-dependent. This study sought to evaluate a novel machine learning software algorithm named VX1 (Volta Medical), trained to adjudicate multipolar electrogram dispersion.

Methods: This study was a prospective, multicentric, nonrandomized study conducted to assess the feasibility of generating VX1 dispersion maps. In 85 patients, 8 centers, and 17 operators, we compared the acute and long-term outcomes after ablation in regions exhibiting dispersion between primary and satellite centers. We also compared outcomes to a control group in which dispersion-guided ablation was performed visually by trained operators.

Results: The study population included 29% of long-standing persistent AF. AF termination occurred in 92% and 83% of the patients in primary and satellite centers, respectively, p = 0.31. The average rate of freedom from documented AF, with or without antiarrhythmic drugs (AADs), was 86% after a single procedure, and 89% after an average of 1.3 procedures per patient (p = 0.4). The rate of freedom from any documented atrial arrhythmia, with or without AADs, was 54% and 73% after a single or an average of 1.3 procedures per patient, respectively (p < 0.001). No statistically significant differences between outcomes of the primary versus satellite centers were observed for one (p = 0.8) or multiple procedures (p = 0.4), or between outcomes of the entire study population versus the control group (p > 0.2). Interestingly, intraprocedural AF termination and type of recurrent arrhythmia (i.e., AF vs. AT) appear to be predictors of the subsequent clinical course.

Conclusion: VX1, an expertise-based artificial intelligence software solution, allowed for robust center-to-center standardization of acute and long-term ablation outcomes after electrogram-based ablation.

Keywords: artificial intelligence; atrial fibrillation; catheter ablation; dispersion; driver; mapping; sinus rhythm.

© The Authors. Journal of Cardiovascular Electrophysiology published by Wiley Periodicals LLC.

Figures

Figure 1

(A) The VX1 software is based on machine learning classification algorithms including a deep learning algorithm. VX1 is trained offline on a large proprietary database. Electrograms are digitized and processed in real‐time by VX1 which in turn provides operators with visual and audio cues representing areas of interest. A direct connection from the acquisition system to a computer installed with the VX1 software allows for data transmission. (B) Flowchart diagram explaining the main steps carried out when the trained algorithm is used on new data. Electrogram information provided by the electrophysiology recording system is processed by a feature extraction module extracting 65 features per single track (Step 1), these features are analyzed by a trained machine learning algorithm to produce an array of dispersion likelihood A (Step 2). In parallel, it is processed by a trained deep learning algorithm (Step 3) producing an array of dispersion likelihood B (Step 4). Both dispersion likelihood arrays are merged using a weighted average based on their agreement level to produce dispersion likelihood C (Step 5). Array C is being color‐coded and displayed on the catheter schematic (Step 6). To account for time‐wise stability, several iterations in time are used to build a different color‐coded list which is displayed on the upper frame of the software interface (Step 7). (C) Schematic description of the tailored ablation protocol implemented in Ev‐AIFib. After biatrial mapping of dispersion regions (Step 1), ablation at these regions was conducted (Step 2). Endpoint: Sinus rhythm  conversion. The ablation set was completed with the connection of the regions ablated/isolated (“Ablate & Connect” or “Circle & Connect”) and in 42% of the patients with Tailored PVI (Step 3).

Figure 2

Kaplan–Meier estimates, for all patients, after a single versus multiple procedures, of (A) freedom from documented atrial fibrillation (p = 0.4), (B) freedom from any atrial arrhythmia (p < 0.001), with or without the use of antiarrhythmic medications. A p < 0.05 was considered statistically significant.

Figure 3

Kaplan–Meier estimates, for primary (blue) versus satellite centers (red), of freedom from any atrial arrhythmia (A) after a single procedure (p = 0.8), (B) after one or more procedures (p = 0.4), with or without the use of antiarrhythmic medications. A p < 0.05 was considered statistically significant.

Figure 4

Kaplan–Meier estimates of freedom from any documented atrial arrhythmia after a single procedure, with or without the use of antiarrhythmic medications, for different groups of patients depending on acute outcomes during the index procedure: (A) patients for which AF was terminated by ablation have a statistically significant lower rate of recurrences than patients for which AF could not be terminated (p = 0.004). (B) Patients converted into sinus rhythm by ablation have a statistically significant lower rate of recurrences than patients converted into atrial tachycardia without further restoring sinus rhythm by ablation (p = 0.03) and patients for which AF could not be terminated (p < 0.001). A p < 0.05 was considered statistically significant. AF, atrial fibrillation.

Figure 5

Kaplan–Meier estimates of (A) freedom from documented AF and (B) freedom from any atrial arrhythmia, after a single procedure for the study group (in blue) versus the control group from the Substrate‐HD study where dispersion was visual (in red). Both visual and VX1‐based dispersion groups are comparable (p = 0.2 and p = 0.9, respectively). A p < 0.05 was considered statistically significant. AF, atrial fibrillation; VX1, Volta Medical.