Multi-parametric Analysis for Atrial Fibrillation Classification in the ECG

Abstract

This study participates in the PhysioNet/CinC Challenge 2017 dedicated to the discrimination of atrial fibrillation (AF) from Normal sinus rhythm (Normal), other arrhythmia (Other) and strong noise using single short ECG lead recordings. Our Matlab entry applies multi-parametric AF classification based on: noise detection; heart rate variability analysis (HRV); beat morphology analysis after robust synthesis of an average beat and delineation of P, QRS, T waves; detection of atrial activity by the presence of a P-wave in the average beat and f-waves during TQ intervals. A Linear discriminant classifier is optimized by maximization of the Challenge F1 score, adjusting the prior probabilities of 4 classes and stepwise selection of a non-redundant feature set. Top-5 features, which contribute to >90% of F1 score are 3 HRV features, P-wave presence and mean correlation of all beats against the average beat. On the blinded test set, our entry has F1 score: 0.89 (Normal), 0.85 (AF), 0.67 (Other), 0.80 (Overall).