Atrial fibrillation classification from a short single lead ECG recording using hierarchical classifier

Atrial fibrillation (AF), one of the most common cardiac arrhythmias, can be diagnosed using electrocardiography. We present a data-driven model to automatically detect the occurrence of atrial fibrillation on a single lead electrocardiogram (ECG). Our model incorporates a wide range of features including heart rate variability in the time and frequency domain, spectral power analysis and statistical modeling of atrial activity. We use an over-sampling strategy to balance the dataset across different categories. We design a hierarchical classification model to predict an ECG signal as either AF, normal, noisy or an alternative rhythm. The best performance was achieved with a hierarchical bagged ensemble classifier, with an average F score of0.7855 over all samples.