Classification of ECG using Ensemble of Residual CNNs with Attention Mechanism

Petr Nejedly1, Adam Ivora2, Ivo Viscor2, Zuzana Koscova2, Radovan Smisek3, Pavel Jurak2, Filip Plesinger2
1Institute of Scientific Instruments of the Czech Academy of Science, 2Institute of Scientific Instruments of the CAS, 3Brno University of Technology, Faculty of Electrical Engineering and Communication, Department of Biomedical Engineering


Here in this paper, we introduce a solution to the Phys-ioNet Challenge 2021. The method is based on the ResNetdeep neural network architecture with a multi-head atten-tion mechanism for ECG classification into 26 indepen-dent groups. The model is optimized using a mixture ofloss functions, i.e., binary cross-entropy, custom challengescore loss function, and sparsity loss function. Probabilitythresholds for each classification class are estimated us-ing the evolutionary optimization method. The final modelconsists of three submodels forming a majority voting clas-sification ensemble. The proposed method can classifyECGs with a variable number of leads, e.g., 12-lead, 6-lead, 3-lead, and 2-lead. The algorithm was trained andvalidated on the public dataset proposed for the challenge.The trained algorithm was tested using a hidden validationset during the official phase of the challenge and obtainedvalidation scores (ISIBrno-AIMT): 0.64, 0.62, 0.63, 0.63,and 0.62 for lead configurations 12, 6, 3, 4, and 2, respec-tively. The total training time was approximately 27 hours,i.e., 9 hours per model