Aims: This study aims to assess the accuracy of a variety of neural network techniques for classification of arrhythmias in electrocardiograms and understand the models.
Methods: We trained a range of neural network architectures on the Physionet Challenge 12-lead ECG database. We trained convolutional networks with max pooling fully convolutional networks, residual networks, as well as LSTM fully convolutional networks. We tuned parameters with cross-validation before evaluating test performance on held-out data.
Results: Overall, a simple convolutional model with max pooling performed best. Interestingly, we found that, for the same 3-layer CNN model, downsampling signals from 500Hz (held out test set F2: 0.608, G2:0.411) to 125Hz (held out test set F2: 0.663, G2: 0.449) improved performance (potentially by reducing the chance to overfit high-frequency noise). We also performed a local feature attribution analysis with Expected Gradients and found that while attributions are effective at capturing small, important sections of a signal like the QRS complex, they often fail to highlight the importance of higher-level features like heart rate or PR interval.
Conclusion: We found that the best predictions of arrhythmia in electrocardiograms came from using a convolutional architecture with downsampled signals. We also found that existing feature attribution methods had only limited success in explaining predictions in this data.