Multi-Label Cardiac Abnormalities Classification on Selected Leads of ECG Signals

Abstract

As part of the PhysioNet/Computing in Cardiology Challenge 2021, Our team, HeartBeats, developed an ensembled model based on SE-ResNet for identifying 30 kinds of cardiac abnormalities from different lead combinations of electrocardiograms (ECGs). At pre-processing stage, ECGs were down-sampled to 500 Hz and each record is normalized using Z-Score normalization. We then employed several residual neural network modules with squeeze-and-excitation blocks to learn from the first 15-second segments of the signals. We designed a multi-label loss to emphasize the impact of wrong predictions during training. We relabelled the dataset which contains only 9 classes using our baseline model build in last year's challenge. Five-fold cross-validation was used to assess the performance of our models. Our classifiers received the scores of 0.58, 0.55, 0.56, 0.53, and 0.53 for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions with the Challenge evaluation metric. Our final model performed well on the test data. However, the results were not officially ranked because our training code may select the offline pre-trained models rather than using the training data if the pre-trained models performed better than the trained models on the training data. The model can therefore fail to learn from new training data.