Application of Deep Learning for Quality Assessment of Atrial Fibrillation ECG Recordings

Abstract

In the last years, atrial fibrillation (AF) has become one of the most remarkable health problems in the developed world. This arrhythmia is associated with an increased risk of cardiovascular events, being its early detection an unresolved challenge. To palliate this issue, long-term wearable electrocardiogram (ECG) recording systems are used, because most of AF episodes are asymptomatic and very short in their initial stages. Unfortunately, portable equipments are very susceptible to be contaminated with different kind of noises, since they work in highly dynamics and ever-changing environments. Within this scenario, the correct identification of free-noise ECG segments results critical for an accurate and robust AF detection. Hence, this work presents a deep learning-based algorithm to identify high-quality intervals in single-lead ECG recordings obtained from patients with paroxysmal AF. The obtained results have provided a remarkable ability to classify between high- and low-quality ECG segments about 92%, only misclassifying around 7% of clean AF intervals as noisy segments. These outcomes have overcome most previous ECG quality assessment algorithms also dealing with AF signals by more than 20%.