Multi Label Multi Lead ECG Classification with Augmenting Convolutional Networks

Tom Beer and Chen Beer


Early diagnosis of cardiac abnormalities from ECG recordings is crucial for predicting and preventing cardiovascular diseases and other morbidities. This laboursome task has not yet been fully automated and remains a heavy burden for cardiologists. As part of the 2020 PhysioNet/Computing in Cardi-ology Challenge, team Karma Backprops developed a convolutional neural network approach for classifying a wide range of cardiac abnormalities from multi lead ECGs. The algorithm is inspired by WaveNet, a system that was proposed origi-nally for the audio domain. This architecture contains various dilation fac-tors that allow its receptive field to grow exponentially with depth. In addi-tion, complex data augmentation mechanisms are utilized, allowing for in-creased generalizability and transportability of the algorithm to unseen envi-ronments. Preliminary evaluation has shown that the proposed approach is capable of correctly diagnosing 9 distinct cardiac events with performance of 0.464 on a hidden test set, where performance is measured by a geometric mean between an F-beta score and a generalization of the Jaccard measure. These encouraging results prove the potential of our method and call for further model development using data from more sources.