Aims: This work provides a viable solution for the automatic identification of cardiac diseases from electrocardiographic recordings (ECG) in the scope of the Computing in Cardiology Challenge. For this purpose we employed a convolutional neural network (CNN), a type of deep learning architecture which is able to learn salient features of input signals in a data-driven manner.
Methods: This work employs a CNN architecture, comprised of convolutional, max pooling, global average pooling and fully connected operations. Batch-normalization and dropout (normal and spatial for fully connected and convolutional operations, respectively) was used as regularization alongside LeakyReLU non-linearities. The employed architecture is depicted in Figure 1. An ECG-tailored data augmentation consisting of white Gaussian noise, powerline noise (50 Hz), baseline wander (0.5 Hz) and voltage offset (1% of signal range) was applied in a “select-one” manner. The challenge database, comprising 6877 ECGs of 9 categories, was split into train and validation sets (75-25%) and windowed to segments of 18432 samples, edge-padding when necessary. The network was trained using stochastic gradient descent (lr = 0.1, momentum = 0.0001, lr reduction on plateau) with a batch size of 128 elements.
Results: The network obtained a F2 score of 74.2% and a G2 score of 53.5% in the validation set and a F2 score of 74.7%, a G2 score of 53.0% (geometric mean of 62.9%) in the independent test set, evidencing the viability of this approach as an initial solution for cardiac abnormality identification.