Analyzing electrocardiogram signals obtained from a nymi band to detect atrial fibrillation

Abstract

In this paper, we propose a method for detecting atrial fibrillation (AF) from electrocardiogram (ECG) signals obtained from a wearable device. The proposed method uses three classification methods: neural networks (NNs), k-nearest neighbors (kNN), and decision trees (DT). The results from each of the three classifiers are combined using a voting system to make the final decision as to whether AF is present. To develop the classification system, we collected data from 61 subjects using a Nymi Band that is wrist-worn ECG monitoring device. From these signals, we extracted the root-mean square of the successive differences (RMSSD) and the Shannon entropy (ShE) of the RR interval, QS interval, and R peak amplitude. These properties were then used as features to train the classifiers. The accuracy, sensitivity, specificity, and precision of this classifier were 97.94%, 100.00%, 96.72%, and 94.74%, respectively for dataset with six features. The ensemble method of NNs, kNN, and DT was evaluated. Depending on the rules for ensemble, the accuracy, sensitivity, specificity, and precision are different among those classifiers. With a rule of unanimous determination for AF, false positive is decreased and false negative is increased. With a rule of unanimous determination for NSR, false positive is increased and false negative decreased. Even though accuracies of each classifier are depending on the set of features, with ensemble method, the accuracy of AF detection can be preserved.