Early detection of Atrial Fibrillation (AF) is essential for preventing heart failure, thrombosis, and cardio-embolic stroke. Traditional neural network (NN)-based detection methods primarily rely on clinical expert diagnoses, yet fully leveraging the intricate patterns inherent in AF episodes remains challenging. Herein, we propose an innovative framework that utilizes statistical inference and probabilistic modeling to analyze cardiac inter-beat interval dynamics, incorporating 5 robust features: heart rate, Differential entropy, minimal redundancy in non-AF episodes, and maximal entropy in AF episodes. The mRMEBP, a Back Propagation NN refined through rigorous optimization, accurately identifies subtle AF patterns. To enhance generalization, we utilize a diverse training set from 4 publicly accessible electrocardiogram (ECG) databases, featuring expert clinical annotations categorized into 6 types using a Fuzzy logic system. Validation against clinical data demonstrates the efficacy of mRMEBP, surpassing state-of-the-art techniques and highlighting its practical value in clinical and healthcare settings.
Read full abstract