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

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Atrial fibrillation (AF) and heart failure (HF) frequently coexist, with AF often preceding HF decompensation.
• Daily remote monitoring of implantable cardioverter-defibrillator (ICD) parameters offers a potential data source for predicting AF.
• Real-world data analysis is crucial for understanding the clinical utility of predictive algorithms.

Purpose of the Study:

• To evaluate the efficacy of machine learning techniques in predicting AF occurrence using daily ICD remote monitoring data.
• To assess the predictive performance of an XGBoost model for AF detection within a 3-day horizon.
• To identify key ICD parameters predictive of future AF events in a real-world patient cohort.

Main Methods:

• Extracted daily ICD transmission data from patients with primary prevention ICDs (2012-2021).
• Trained an XGBoost model to predict AF occurrence within 3 days, using a 14-day data sequence.
• Validated model performance retrospectively and prospectively using ROC AUC and PR AUC; assessed feature importance with SHAP values.

Main Results:

• The XGBoost model achieved an AUROC of 0.79 and AUPRC of 0.10 for predicting AF within 3 days.
• The model demonstrated high specificity (99%) in the validation dataset.
• Key predictive variables included right ventricular (RV) and right atrial (RA) sensing amplitudes and pulse width.

Conclusions:

• Machine learning applied to daily ICD remote monitoring data shows potential for predicting AF occurrence.
• The developed risk prediction algorithm requires external validation in large-scale, multi-center clinical trials.
• This approach may aid in the early detection and management of AF in patients with coexisting HF.