Machine learning-lased prediction of atrial fibrillation in patients with atrial high-rate episodes

Atrial high-rate episodes (AHREs) detected by cardiac implantable electronic devices (CIEDs) are associated with an increased risk of atrial fibrillation (AF), stroke, and mortality. Several predictive models, including CHA₂DS₂-VASc, C₂HEST, mCHEST, HAT₂CH₂, and HAVOC, estimate AF risk in patients with CIEDs but rely on conventional statistical methods and only perform modestly for prediction. These clinical risk models do not incorporate machine learning techniques, which could improve predictive accuracy by capturing complex risk factor interactions. This study addresses this gap by applying novel machine learning (ML) algorithms to enhance AF risk prediction.

This study aims to predict AF using novel ML techniques in patients with at least one AHRE detected by CIEDs.

This prospective longitudinal study included patients with at least one AHRE detected by CIEDs and no prior history of AF. Machine learning models were applied to predict AF. To handle missing data, the Multivariate Imputation by Chained Equations method was used, ensuring no significant bias through chi-square testing. Class imbalance was addressed with Random Over-Sampling Examples to enhance predictive performance. Various ML algorithms, including CatBoost, Random Forest, AdaBoost, and Naive Bayes, were evaluated for their predictive accuracy. Data analysis was conducted using R (version 4.3.1).

A total of 100 patients (47% male, mean age 65.98 ± 18.01 years) were included in the final analysis, of whom 24 developed AF (24%) after 1-year follow-up. Among implanted cardiac devices, 87% had a permanent pacemaker, 4% an implantable cardioverter-defibrillator, and 9% cardiac resynchronization therapy. A CHA₂DS₂-VASc score of 0, 1, and 2 was observed in 6%, 23%, and 18% of patients, respectively, while 53% had a score of ≥3. Mild, moderate, and severe left atrial enlargement was seen in 29%, 6%, and 2% of patients, respectively. Among the ML models evaluated, the CatBoost algorithm demonstrated the highest predictive performance, achieving an F1-score of 0.789 and an AUC of 0.896 (95%CI (0.768 – 1.000), with an overall accuracy of 0.724. While Random Forest and AdaBoost achieved the highest accuracy (0.759), they had slightly lower AUC values (0.851 (95%CI 0.741 - 1.000) and 0.792 (95%CI 0.628 - 0.957), respectively. Naive Bayes demonstrated the highest sensitivity (0.818) but had lower specificity (0.571). The top three most influential predictors of AF in the CatBoost model were hypertension, diabetes, and age.

ML techniques are effective in predicting AF in patients with AHREs. These models offer valuable insights and may enhance early risk stratification in clinical practice. Further studies with larger cohorts and extended follow-up periods will validate their predictive power and optimize their clinical application.

Evaluation Metrics