Predicting Atrial Fibrillation Ablation Outcomes: Machine Learning Model Development and Validation Using a Large...

Area of Science:

• Cardiology
• Health Informatics
• Machine Learning in Healthcare

Background:

• Atrial fibrillation (AF) ablation is effective but has inconsistent long-term success rates (50-75%).
• Current patient selection relies on individual assessment, lacking population-level predictive analytics.
• Existing risk scores (CHADS₂, CHA₂DS₂-VASc, CAAP-AF) have unclear performance in administrative claims data.

Purpose of the Study:

• To evaluate if machine learning (ML) models integrating International Classification of Diseases (ICD) billing codes outperform traditional risk scores in predicting 1-year AF ablation outcomes.
• To explore the utility of large administrative claims databases for developing standardized, scalable prediction models for AF ablation.

Main Methods:

• Analysis of 14,521 patients undergoing AF ablation from Medicare claims data (2013-2020).
• Development of logistic regression and extreme gradient boosting (XGBoost) models using demographics, comorbidities, and ICD codes from the 2 years prior to ablation.
• Comparison of ML model predictions against claims-based CHADS₂, CHA₂DS₂-VASc, and modified CAAP-AF risk scores.

Main Results:

• AF ablation success was observed in 54.01% of patients.
• XGBoost models demonstrated better discrimination (AUC 0.528-0.529) than traditional risk scores (AUC ~0.50) across all groups.
• Models incorporating ICD codes showed improved performance, especially in paroxysmal AF patients (AUC 0.544).

Conclusions:

• ML models using administrative data offer statistically significant but modest improvements over traditional risk scores for predicting AF ablation outcomes.
• The predictive performance is currently insufficient for clinical decision-making due to limitations in administrative data (lack of clinical variables).
• Findings support the use of these models for health system quality monitoring and comparative effectiveness research, rather than individual patient management.