Selected Features for Classification of 12-lead ECGs

Abstract

In this paper we describe our algorithm develop by the Alba_W.O. team at The PhysioNet/Computing in Cardiology Challenge 2020. Our approach achieved a challenge validation score of 0.308 and a full test score of 0.102, placing us 31 out of 40 in the official ranking. Our final algorithm is based on bootstrap-aggregated (bagged) decision trees. For the classification task, we provide a set of features extracted from 12-lead ECG, in detail describe later. We use the method implemented in PhysioNet-Cardiovascular-Signal-Toolbox: Global Electrical Heterogeneity, arterial fibrillation features, and PVC detection. We also estimate ECG periods (PR, QS, QR, PT, TP) and morphology parameters (ST elevation, QRS area, ECG value at R points). We also examine the importance of each predictor individually, for the classification task, using a t-test. All groups of used parameters, without sex shown utility in some class classification cases.