Cardiac rhythm classification from a short single lead ECG recording via random forest

Abstract

Detection of atrial fibrillation (AF) from electrocardiogram (ECG) recordings is one of the prevailing challenges in the field of cardiac computing. The task of the PhysioNet/Computing in Cardiology 2017 challenge is to distinguish the AF rhythms from non-AF rhythms using a short single lead ECG recording. In this study, we analyzed 62 time and frequency-domain, linear, and nonlinear features to discriminate four classes, viz., normal sinus rhythm, AF, noisy, or other rhythm. The feature space dimension was reduced to 37 using a Genetic Algorithm based feature selection. We trained a random forest classifier on the given 8,528 training dataset and obtained a ten-fold cross validation classification accuracy of 82.7%. On the test dataset, we obtained an F1-score of 0.91, 0.74, and 0.70 for NSR, AF, and other rhythms, respectively. Results suggest that with the proposed model it is possible to classify cardiac abnormalities from a single lead ECG even when the recordings are of short duration.