Automatic detection of atrial fibrillation and other arrhythmias in holter ECG recordings using rhythm features and neural networks

Abstract

Atrial fibrillation (AF) is a disease affecting 1-2 % of the population. Due to its episodic behavior, it is usually detected using Holter recordings. While various AF detection methods have been described in the past, it still remains problematic because holter recordings may contain other arrhythmias (OA) and, moreover, they may be influenced by patient movements. In accordance with the Physionet Challenge 2017, we propose an autonomous and robust method of distinguishing between pathological and normal recordings. First, QRS complexes are identified using envelograms (8-25 Hz and 70-90 Hz). Detected QRS complexes are clustered into morphology groups using a raw ECG signal. If too many morphology groups are produced or less than four QRS complexes are detected, the process is aborted and the recording is considered too noisy. Next, a median shape for the first and the second morphological group of QRS complexes is built. Features are extracted from averaged shapes, from the rhythm of major morphology QRS complexes, from QRS correlation to template shapes and from the convolutional neural network. 277 features are fed into the neural network, resulting in three outputs. The 120 most important features, as well as outputs from the neural network, are fed into a bagged tree ensemble. Machine-learning algorithms and logical rules were trained using 8,138 files from a reduced training set. The resultant F1 score measured using hidden test set (3,658 recordings) was 0.81 (normal 0.91, AF 0.80, OA 0.74).