Cardiac Abnormalities Recognition in ECG Using a Convolutional Network with Attention and Input with an Adaptable Number of Leads.

In this work, we present an algorithm for automatically identifying the cardiac abnormalities in ECG records with the various number of leads. The algorithm is based on the modified ResNet convolutional neural network with the attention layer. The network input is modified to allow using a single network for different lead subsets. In an official phase challenge entry, our BUTTeam reached the 15th place. In our test challenge entry, we have achieved 0.470, 0.460, 0.470, 0.460, and 0.460 of the challenge metric for 12, 6, 4, 3 and 2 leads with ranking 14th, 14th, 11th, 15th and 11th place, respectively. From additional evaluation of other lead subsets, the leads representing a common heart axis orientation achieved the best detection results. However, all lead subsets performed very similarly.