Classification of ECG Using Ensemble of Residual CNNs with Attention Mechanism

Abstract

This paper introduces a winning solution (team ISIBrno-AIMT) to the PhysioNet Challenge 2021. The method is based on the ResNet deep neural network architecture with a multi-head attention mechanism for ECG classification into 26 independent groups. The model is optimized using a mixture of loss functions, i.e., binary cross-entropy, custom challenge score loss function, and sparsity loss function. Probability thresholds for each classification class are estimated using the evolutionary optimization method. The final model consists of three submodels forming a majority voting classification ensemble. The proposed method classifies ECGs with a variable number of leads, e.g., 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead. The algorithm was validated and tested on the external hidden datasets (CPSC, G12EC, undisclosed set, UMich), achieving a challenge score 0.58 for all tested lead configurations. The total training time was approximately 27 hours, i.e., 9 hours per model. The presented solution was ranked first across all 39 teams in all categories.