Combining template-based and feature-based classification to detect atrial fibrillation from a short single lead ECG recording

Automated diagnosis of Atrial fibrillation (AF) has remained imperfect despite the threat it represents to millions of people. The main issues which can lead to a misdiagnosis of AF include its episodic nature, disease diversity and noise. The aim of 2017 PhysioNet/CinC Challenge is to classify short single lead ECG recordings as normal sinus rhythm, atrial fibrillation, other rhythm, or noisy. We present a method using heart rate variability features and noise detection features coupled with template-based wave morphology features. The method originality lies in the use of special templates sensitive to the heart rate variability as well as wave morphology. These special templates showed significant results in AF detection performances. Based on Cross-validation, an F1 score of 0.84 on AF classification, and a general classification score of 0.76 were obtained on the training set.