Classification of atrial fibrillation in short-term ECG recordings using a machine learning approach and hybrid QRS detection

Atrial fibrillation (AF) is one of the most common sustained cardiac arrhythmia, occurring in 1–2% of the general population. Significant mortality and morbidity is related to occurrence of AF arrhythmia due to high risk of hospitalization, stroke, heart failure and coronary artery disease, etc. In many cases AF may not produce any symptoms and may go unnoticed by a patient, which is why there is a high importance to develop methods of detecting this heart disorder. Creating an algorithm for AF and other arrhythmias classification of short-term single lead ECG signals was the aim of the PhysioNet Challenge 2017. The database was composed by over 8.5 thousand ECG recordings (between 10 sec and 60 sec length) measured by AliveCor device, provided by organizers. We prepared an alternative hybrid approach for QRS detection in order to obtain RR time intervals. It consists of two complementary methods in hierarchical order: one based on nonlinear transformation and first-order Gaussian differentiator as superior and another one proposed in sample entry as inferior. We introduce the machine learning algorithm in order to classify whether it is normal sinus rhythm, AF or an alternative heart rhythm using features considered regularity of RR time intervals and morphology of the ECG signal. The separate part of the algorithm based on beat averaging method is dedicated for preceding extraction of too noisy recordings from the input to the classifier. The best overall F1 score we achieved in the official phase of the PhysioNet Challenge 2017 was 0.77 (0.86 for normal, 0.78 for AF and 0.66 for other rhythms).