A Machine Learning Approach for Atrial Fibrillation Detection in Telemonitored Patients

Abstract

Atrial fibrillation (AF) is the most common type of cardiac arrhythmia. As it is typically asymptomatic, it often goes undiagnosed until major complications arise, such as stroke. Therefore, the development of rapid, economical, and widely accessible diagnostic tools for detecting AF at an early stage is crucial. Telemonitoring with machine learning-assisted devices shows promise in achieving this goal. This paper presents an algorithm that automatically detects AF in signals obtained by portable electrocardiographs connected to a telemonitoring platform via smartphones. The algorithm consists of three stages: a noise detection, ectopic beat removal and an AF detection. The noise detection involves analyzing the ECG signals using 5-s windows with a 1-s shift. A K-nearest neighbors (KNN) classifier predicts the presence or absence of noise in each window, allowing for the detection of noisy and non-noisy segments of the signal. The non-noisy segments are processed using a Pan-Tompkins algorithm to find the R peaks of the signal, and the corresponding RR interval series. Then ectopic beats are removed using an XGBoost classifier, generating the NN series. In the AF detection stage, X features are obtained from this series, which serve as input features of an XGBoost classifier that predicts the presence or absence of AF in the ECG signal. The algorithm was trained and tested using the Physionet Short Single-Lead AF Database (SSLAFDB) and achieved an accuracy of 90.87% and an F1-score of 90.91%. Further validation was performed by an external partner using two other databases, reporting an accuracy of 90.41% and 89.61% respectively.