Detection of atrial fibrillation episodes from short single lead recordings by means of ensemble learning

Abstract

An approach is presented to classify ECG signals as normal, atrial fibrillation, other arrhythmia, or noisy in the context of the Physionet/CinC challenge 2017. The presented approach is a two-stage one, where first noisy recordings are detected based on generic features in the data. Then in the second stage known indices for atrial fibrillation are used as features. For both stages an ensemble model with decision trees is used, fitted with RUSBoost to account for the class imbalance in the dataset. With this approach an overall F1 score of 0.75 is obtained. The method achieves an accurate classification of AF signals, but the misclassification for other arrhythmia is relatively high. Suggestions are also presented on how ECG wave morphologies could be taken into account by using deep learning to further improve the classification.