ArNet-ECG: Deep Learning for the Detection of Atrial Fibrillation from the Raw Electrocardiogram

Noam Ben-Moshe1, Shany Biton2, Joachim A. Behar3
1Faculty of Computer Science, Technion-IIT, 2Faculty of Biomedical Engineering, Technion-IIT, 3Technion-IIT


Introduction: Atrial fibrillation (AF) is the most prevalent heart arrhythmia. We hypothesize that using a deep learning (DL) approach trained on the raw electrocardiogram (ECG) will enable robust detection of AF events and the estimation of the AF burden (AFB). This is due to the characteristics of AF which includes irregular beat-to-beat time interval variation, the absence of P-wave and the presence of fibrillatory waves (f-wave). We develop a new DL algorithm, denoted ArNet-ECG, to robustly detect AF events and estimate the AFB from the raw ECG.

Methods: A dataset including 2,247 adult patients and totaling over 53,753 hours of continuous ECG from the University of Virginia (UVAF) was used.

Results: ArNet-ECG obtained an F1-score of 0.94. The median absolute AFB estimation error was |EAF(%)| = 0.35%.

Discussion and conclusion: We created a new model called ArNet-ECG that uses the raw ECG as input to detect AF events. Future work includes the integration of a recurrent neural network unit to take into account the time dependency between consecutive events, assessing the generalization of ArNet-ECG on external test sets of different ethnic groups. Finally, ArNet-ECG needs to be benchmarked against state-of-the-art models taking the beat-to-beat intervals as input.