Abstract
Background: Atrial fibrillation (AF) ablation is an effective treatment for reducing episodes and improving quality of life in patients with AF. However, in some patients there are only modest long-term AF-free rates after AF ablation. There is a need to address the limited benefits some patients experience by developing predictive algorithms to improve AF ablation outcomes. Objective: The authors aim to utilize machine learning models on claims data to explore if innovative coding models may lead to better patient outcomes than use of traditional stroke risk score prediction. Methods: The Merative MarketScan® Research Medicare data was used to examine claims for AF ablation. To predict 1-year AF-free outcomes after AF ablation, logistic regression and XGBoost models were used. Model predictions were compared with established risk scores CHADS 2 and CHA 2 DS 2 -VASC. These models were also assessed on subgroups of patients with paroxysmal AF, persistent AF, and both AF and atrial flutter from October 2015 onwards. Results: The sample included 14,521 patients with claims for AF ablation. XGBoost achieved an area under the receiver operating characteristic curve (AUC) of 0.525, 0.521, and 0.527 for the entire AF ablation population, female, and male, respectively. Within the subgroups, machine learning models performed the best for the paroxysmal AF subgroup using ICD codes, demographic information, and comorbidity indexes, achieving an AUC of 0.546. Conclusion: Machine learning models outperformed CHADS 2 and CHA 2 DS 2 -VASC in all AF ablation patient groups (whole population, female, and male). Using patient data for those who had their AF ablation on or after October 2015, machine learning models performed best in all subgroups and the population, indicating that including ICD codes in machine learning models may improve performance.

This publication has 27 references indexed in Scilit: