Deep Learning Approaches for Predicting Atrial Fibrillation Risk in Pacemaker Patients

Pacemaker wearers represent a vulnerable population, often facing cardiac issues and a heightened risk of cardiovascular diseases, such as Atrial Fibrillation (AF). Timely identification of patients at risk of developing AF is of paramount importance in preventing potentially serious complications, including strokes, limb ischemia, and other cardioembolic consequences. Deep Learning (DL) methods have demonstrated significant efficacy in AF prediction, offering a promising avenue to enhance the management of this condition through electrocardiogram (ECG) analysis. To overcome data scarcity challenges, we have embraced MLOps, a cutting-edge methodology. It offers a systematic framework to manage DL model lifecycles, ensuring robustness and reproducibility from development to deployment and monitoring in our research. However, the scarcity of digitized and labeled data prompted us to explore data generation techniques, including conditional generation and standard augmentation, subsequently subjected to meticulous evaluation. Assessment of diverse data generation methods reveals that ECGs synthesized via conditional DeepFake generation exhibit an improved distribution, characterized by MMD scores = 8.10-3 and MMD = 4.10-4. This consequently demonstrates acceptable FD, RMSE, and PRD scores relative to a baseline. A series of transfer learning efforts culminated in relatively conclusive outcomes, illustrated by a 98% F1- score and an AUC of 0.97, leveraging a larger dataset. Despite inherent limitations related to data size, this study, fortified by its rigorous methodology, highlights the efficacy of transfer learning, introduces a pertinent comparative approach to data generation techniques, and facilitates the assembly of a dedicated ECG dataset sourced from patients equipped with pacemakers.