Multi-label Cardiac Abnormality Classification from Electrocardiogram using Deep Convolutional Neural Networks

Nima Wickramasinghe1 and Mohamed Athif2
1Department of Electronic and Telecommunication Engineering, University of Moratuwa, 2Department of Biomedical Engineering, Boston University


Introduction: The proposed method uses a deep neural network architecture to perform multi-label classification of 26 cardiac abnormalities from 12-lead and reduced lead ECG data. The model was created by team “NIMA” for the PhysioNet/Computing in Cardiology Challenge 2021. Methods: ECG signals of length at most 20 seconds (97% of the dataset) were used for training. We preprocess the data by resampling the signals at 100Hz and zero-padding to 2000 samples. The preprocessed ECG signals and Fast Fourier Transforms obtained from the preprocessed signals are each fed into two separate deep Convolutional Neural Networks. Each of the networks has a depth of four convolutional layers that use the Swish activation function. We use spatial dropouts and average pooling between each convolutional layer to reduce overfitting and reduce model complexity. Following the four convolutional layers, the time and frequency domain network outputs are concatenated to form a vector of size 1152. The final stage of the model has two dense layers of size 1152 and 26 that use Swish and Sigmoid activation functions, respectively. A threshold of 0.3 is used on the output probabilities to determine the class to address data imbalance. We trained the model for approximately 20 epochs with 79744 training samples per epoch, using the Adam optimizer. The model was implemented using Python Tensorflow using the Keras functional API. Results: The method achieved a score of 0.609, 0.579, 0.595, 0.589, and 0.584 on 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead validation datasets, respectively, according to the PhysioNet/CinC 2021 Challenge scoring metric. Conclusion: Our model produces results of comparable performance to 12 Lead ECG using fewer leads through the use of both time and frequency domain signals while addressing the challenge of data imbalance through the choice of network architecture and preprocessing.