Atrial Fibrillation Treatment Stratification Based on Artificial Intelligence-Driven Analysis of the Electrophysiological Complexity

Abstract

Background: Atrial Fibrillation (AF) treatment strategies are suboptimal and clinical predictors of success are limited. Artificial Intelligence (AI) has arisen as a powerful tool for treatment efficacy prediction.

Objective: We developed an AI-driven platform for the stratification of patients based on noninvasive Electrocardiographic Imaging (ECGI) biomarkers and clinical parameters to evaluate and predict optimal patient treatment.

Methods: We evaluated 204 patients treated according to clinical guidelines and characterized them at the electrophysiological level using ECGI recordings during AF. ECGI signals were calculated to obtain frequency and rotational biomarkers. Baseline clinical characteristics and treatment after inclusion were registered.

Results: A clustering algorithm was calibrated taking three different variables for 1 year outcome prediction: (1) AF type (paroxysmal or persistent); (2) ECGI complexity score (calculated based on highest dominant frequency, median dominant frequency, and mean rotor time); and (3) type of treatment: rhythm control (drugs, AF ablation) or rate control. The cluster analysis classified patients into five groups: Low electrophysiological complexity patterns were associated with an improved outcome after ablation, regardless of the time duration of the AF. Intermediate complexity scores in paroxysmal AF had a favourable outcome with rhythm control treatments, but not in persistent AF patients. Cluster patterns with higher electrophysiological complexity were associated with a higher probability of AF recurrence, both in paroxysmal and persistent groups. The performance of the algorithm predicting the outcome was (AUC: 0.73 (0.63-0.81)), increasing overall performance with respect to conventional persistent and paroxysmal classification (AUC: 0.58 (0.48-0.68); p < 0.05). This algorithm was evaluated on the 20% test set, obtaining 90% prediction success.

Conclusions: AI-driven analysis that combined clinical information with ECGI biomarkers increased the performance of conventional classification methods for AF treatment stratification.

Keywords: ECGI; artificial intelligence; atrial fibrillation; stratification.

Figure 1

Sample DF and rotor histogram maps from two selected patients. (A) Patient showing simple and homogeneous distribution of activation frequency associated with low values of the ECGI complexity score with localized rotors in the pulmonary vein (panel C). (B) Patient showing a heterogeneous distribution of activation frequency associated with high values of the complexity score showing rotational activity in the right atrium (panel D).

Figure 2

Evaluation of long‐term treatment efficacy (freedom from AF at 1 year) of patients with AF (from left to right): 1. Clinical predictors of treatment efficacy at 1 year using multivariant analysis; 2. Complexity Score Evaluation: ECGI using feature extraction; 3. Score calculation (Formula1); 4: AF treatment stratification: Clustering algorithm analysis.

Figure 3

Score calculation: Examples of patients belonging to each of the five clusters. Panel describes the AF type, treatment, dominant frequency map, ECGI biomarkers for each specific patient, calculated score, and clinical outcome.

Figure 4

Results obtained from the clustering algorithm identifying 5 different groups in three levels. From left to right, the first variable represents AF type (Paroxysmal on top, Persistent on the bottom), score value divided in three subgroups for the paroxysmal patients and in two groups for the persistent patients and type of treatment and 1‐year outcome in the right part of the diagram. Green and red colours represent AF Freedom and AF at 1 year, respectively.