Artificial intelligence-based prediction of arrhythmia recurrence following atrial fibrillation ablation – a use case of the GAIA-X-MED project

Background

Atrial fibrillation (AF) is the most common cardiac arrhythmia in adults and is associated with a significant morbidity and mortality. Catheter ablation is the cornerstone of the rhythm control strategy in AF and is performed on a large scale worldwide. Despite the significant technological improvements and increased operators experience, the efficacy of catheter ablation in maintaining the sinus rhythm remains limited, especially in patients with persistent AF and important left atrial substrate.

Significant advances in artificial intelligence (AI) have already led to numerous applications in medicine. 

Purpose

To evaluate the feasibility of developing an AI-based algorithm able to predict the risk of arrhythmia recurrence after AF ablation as a cooperation between clinics and AI providers following the prerequisites of the GAIA-X project.

Methods

The environmental dataset comprised 288 patients undergoing catheter ablation for AF by means of three-dimensional electroanatomical mapping (EAM) systems and included in a prospective local registry, encompassing baseline, procedural and follow-up characteristics. The project was a collaboration between a tertiary electrophysiology centre and institutes of telematics, bio- and neuroinformatics and software engineering and programming languages, as well as private partners. 

Ordinal and categorical variables were pre-processed separately, with ordinal variables adjusted by subtracting the mean and imputing empty entries with this mean. Categorical data were transformed into a single coding for each possible feature expression, and missing entries were assigned to a distinct category. The variables were chosen after a thorough literature search of the described AF recurrence predictors using AI (Table 1). 

Results

Data from the 288 patients were included. A multi-layer perceptron (MLP) model was employed, receiving both categorical and ordinal features as input. The categorical feature encoding was converted into learnable embeddings, with the architecture allowing for dynamic adjustments to the number of layers and the size of embeddings, while regularization was applied through dropout. The model was trained using cross-entropy loss with the AdamW optimizer, configured with a learning rate of 0.001 and a weight decay of 0.9. This resulted in the development of a toll receiving the baseline and procedural characteristics as inputs and generating the risk of a post-ablation atrial arrhythmia recurrence (Figure 1). 

Conclusion

The development of an AI-based algorithm to predict the risk of arrhythmic recurrence following AF-ablation as part of the GAIA-X-MED project was feasible in this pilot study. This approach has the possibility to select the patients at increased risk of arrhythmia recurrence, who could benefit from more intensive treatment or a closer follow-up. Larger studies are needed to validate the results and draw definitive conclusions.