Leveraging Period-Specific Variations in ECG Topology for Classification Tasks

Abstract

We explore whether specific time-varying shape characteristics of electrocardiograms can be tapped to inform computational approaches in classifying cardiac abnormalities. In particular, we train a random forest classifier on features derived from relative differences between algebraically-computable topological signatures of consecutive segments within ECGs. We convert segments of ECGs as point cloud embeddings in high-dimensional space, extract their topological summaries, and compare these via statistical descriptors and different metrics. As part of the PhysioNet/Computing in Cardiology Challenge 2021, we (Team Cordi-Ak) test this approach across full-and reduced-lead ECGs. Using the Challenge's evaluation metric, our classifiers received scores of -0.06, -0.07, -0.08, -0.08, and -0.10 (consistently ranked 35th out of 39 official entries) for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions of the hidden test set.