Aim: This study (from Revenger team) aims to develop effective approaches for the detection of cardiac arrhythmias from varying-dimensional electrocardiography (ECG) in the PhysioNet/Computing in Cardiology Challenge 2021, taking advantage of both deep neural networks (DNNs) and insights from clinical diagnostic criteria.
Methods: 26 classes (equivalent classes are counted one) of ECGs are divided into two categories. Detectors are manually designed for classes in the category with clear, easy-to-describe clinical rules. The rest classes with subtle morphological and spectral characteristics are classified by DNNs. Input ECGs are resampled (500Hz), bandpassed, and normalized. To make the networks capable of capturing features of different scopes, we use multi-branch convolutional neural networks (CNNs), each with different receptive fields via dilated convolutions. Considering ECGs’ varying dimensionality, convolutions are grouped with group number equaling the number of leads. Adaptive thresholding is used to give final predictions. Outputs from DNNs and from manual detectors are merged to give final predictions.
Results: The best score (challenge metric) of our approach are 0.51, 0.49, 0.47, 0.48, 0.48 for the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead versions of the hidden validation set respectively.
Conclusion: The proposed hybrid method is effective for establishing auxiliary diagnosis systems, and the reducedlead (reduced from the standard 12-lead) ECGs are sufficient for such systems.
Code, configurations and auxiliary data are available at https://github.com/DeepPSP/cinc2021