Adaptive Lead Weighted ResNet Trained with Different Duration Signals for Classifying 12-lead ECGs

Abstract

Introduction: We describe the creation of a deep neural network architecture to classify cardiac abnormality from 12 lead ECGs. The model was created by the team 'between a ROC and a heart place' for the Phys-ioNet/Computing in Cardiology Challenge 2020. Methods: ECGs were downsampled to 257 Hz and then set to a consistent duration by randomly clipping or zero-padding the signal to 4096 samples. To learn effective features, we created a modified ResNet with larger kernel sizes that models long-term dependencies. We embedded a Squeeze-And-Excitation layer into the modified ResNet to learn the importance of each lead, adaptively. A simple constrained grid-search was applied to deal with class imbalance. Results: Using the bespoke weighted accuracy metric, we achieved a 5-fold cross-validation score of 0.684, sensitivity and specificity of 0.758 and 0.969, respectively. We scored 0.520 on the full test data, and ranked 2nd out of the 41 teams that participated in the challenge. Conclusion: The proposed prediction model performed well on the validation and hidden test data. Such models may be potentially used for ECG screening or diagnosis.