Aim: This study aims to develop robust and high performance deep neural network models, more specifically convolutional recurrent neural networks (CRNNs), for the detection of cardiac arrhythmias from 12-lead electrocardiogram (ECG). This would help improving early detection and noninvasive monitoring of cardiovascular diseases (CVDs).
Methods: 12-lead ECGs are sliced into non-overlapping segments (of typical length 6 seconds). The models mainly consist of two sub-networks. The first part is a convolutional neural sub-network (CNN), which captures ECG features from coarser ones, like regularity of RR (R wave to R wave) intervals, to finer ones, like morphology of certain complex or even subtle shapes of specific waves in certain leads. Afterwards, feature tensors extracted from the CNN are fed into a recurrent neural sub-network (RNN), which is good at handling series data. A typical RNN used is composed by 3 bidirectional long short-time memory (LSTM) layers. Finally, a fully connected layer with sigmoid activation is applied to produce the predictions, which are probabilities of belonging to each type (one normal sinus rhythm type and eight abnormal types). A threshold of 0.5 is used to give the final judgement.
Results: In our experiments, considering the amount of the CPSC2018 data, we adopted VGGNet, simple yet comparably powerful, as the CNN feature extractor. Stratified train-validation-test split with ratio of 8:1:1 was made for training and evaluating models. F_beta and G_beta of the final model were 0.58 and 0.37 resp. on the test set, and these two scores provided by the CINC2020 Challenge submission system were 0.53 and 0.34 resp.
Conclusion: CRNNs offer a powerful and easily expandable framework of detecting cardiac arrhythmias for 12-lead ECGs. In order to better capture both global and local ECG features for improving the overall performance, more careful design of the convolutional sub-network should be made and tested.