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

Given the modest performance of available predictive models in estimating the risk of atrial fibrillation (AF) in patients with atrial high-rate episodes (AHREs) detected by cardiac implantable electronic devices (CIEDs), this study explores the potential use of machine learning (ML) algorithms in this context.

To assess the ability of ML techniques in identifying patients with AHRE at high risk of AF.

In this prospective study, we enrolled patients without a prior history of AF who experienced at least one AHRE episode detected by CIEDs. ML techniques were applied to predict the 1-year risk of developing new-onset AF based on the following variables: age, BMI, sex, smoking, hypertension, diabetes, coronary artery disease, chronic kidney disease, dyslipidaema, history of stroke or transient ischaemic attack, vascular heart disease, left atrial enlargement (LAE) and congestive heart failure.

Study population consists of 100 patients (48% male, mean age 66.0 ± 18.0 years), of whom 24 developed AF (24%) after 1-year follow-up. The CatBoost ML model achieved the highest AUC (.857, 95% CI .671-.999) when compared to other ML models and all clinical risk scores. The top four most influential predictors of AF in the CatBoost model were LAE, hypertension, diabetes and age.

ML techniques are robust in predicting AF in patients with AHREs. Further validation in larger, independent cohorts is warranted.