Leveraging Implantable Cardiac Defibrillator Remote Transmissions to Predict the Occurrence of Atrial Fibrillation.

Abstract

Background: Atrial fibrillation (AF) and heart failure (HF) frequently coexist in patients, with the development of AF often preceding HF decompensation. We sought to evaluate whether daily remote monitoring of ICD parameters could predict AF occurrence using machine learning techniques in a real-world cohort.

Methods: Data from patients with primary prevention ICDs transmitted daily to the Northwell centralized remote monitoring center between 2012 and 2021 were extracted. Using this data, an XGBoost model was trained to predict AF occurrence with a 3-day time horizon using a 14-day data collection sequence. Model predictive performance was validated retrospectively and prospectively, using mean ROC AUC and PR AUC across all folds. Feature importance was assessed using Shapley additive explanation (SHAP) values.

Results: A total of 207 patients, 69.0% male, median age of 65.0 [57, 72] years, median ejection fraction of 30% [25, 40], 13.0% paroxysmal AF, and 35.7% with ischemic cardiomyopathy were monitored for over 36 months. Our model predicted AF occurrence within the following 3 days in 49 (23.7%) patients after a median of 36 months post-implant with an area under the receiver operating characteristic curve (AUROC) of 0.79 and an area under the precision-recall curve of 0.10 (AUPRC). The model has a specificity of 99% in the validation data set. Key variables included RV and RA sensing amplitudes as well as the pulse width. Validation was performed using K-fold cross-validation methods without a significant drop in performance metrics.

Conclusion: This exploratory analysis suggests a machine learning approach has the potential to predict AF from daily remote monitoring of ICD parameters. This risk prediction algorithm requires external validation in a large-scale multi-center clinical trial.