Classification and Analysis of 12-Lead Electrocardiograms

This work investigates the classification of 12-lead electrocardiogram (ECGs) to detect abnormalities in the heart using three computational techniques. They are: (1) gradient-boosted ensembling following manual feature extraction, (2) deep learning with stacked autoencoders connected to the output of a multi-layer perceptron (MLP) classifier, and (3) a fusion model combining deep-learning and manually extracted features. An experiment is conducted using the PhysioNet/Computing in Cardiology Challenge 2020 database, addressing a multi-label classification task involving 27 heartbeat rhythm diagnoses. The best-performing model, which merges handcrafted features with autoencoder-derived features, achieves an average classification accuracy of 30.7% and a challenge metric score of 0.4366. The paper concludes by discussing potential improvements in multi-channel ECG classification methods.