Predicting arrhythmia recurrence post-ablation in atrial fibrillation using explainable machine learning.

Area of Science:

• Cardiology
• Medical Imaging
• Machine Learning

Background:

• Distinguishing patients at risk for arrhythmia recurrence post-atrial fibrillation ablation is clinically challenging.
• Explainable machine learning (xML) offers new methods for assessing recurrence risk.
• Developing predictive models can guide post-ablation follow-up strategies.

Purpose of the Study:

• To develop an xML algorithm predicting arrhythmia recurrence after ablation.
• To identify key patient-specific risk factors for recurrence.
• To enhance clinical decision-making for post-ablation patient management.

Main Methods:

• Reconstructed pre- and post-ablation left atrium models using LGE-MRI from 67 patients.
• Trained a random forest classifier using LGE-derived substrate, LA geometry, and clinical factors.
• Utilized SHapley Additive exPlanations (SHAP) to determine feature importance for recurrence prediction.

Main Results:

• The classifier accurately predicted recurrence (ROC AUC: 0.80, PR AUC: 0.82).
• Key predictors identified: larger left atrium, reduced post-ablation atrial floor scar, and prior cardioversion.
• Model validation in a 15-patient cohort showed 80% accuracy.

Conclusions:

• SHAP-based xML provides a proof-of-concept clinical tool for explaining arrhythmia recurrence risk.
• Combines patient clinical profiles with LGE-MRI data for personalized risk assessment.
• Facilitates tailored follow-up strategies after atrial fibrillation ablation.