Arrhythmia classification of 12-lead Electrocardiograms by Hybrid Scattering-LSTM networks

Philip Warrick1, Masun Nabhan Homsi2, Vincent Lostanlen3, Michael Eikenberg4, Joakim Andén5
1Perigen, 2Simon Bolivar University, 3New York University, 4Flatiron Institute, 5KTH Royal Institute of Technology


Abstract

Aims: Electrocardiogram (ECG) analysis is the standard of care for the diagnosis of cardiac abnormalities. Automatic analysis of ECG could aid cardiologists to make more accurate predictions of different types of arrythmias, for timely treatment and reduced healthcare costs. We developed an arrhythmia detector of 12-channel ECG using Scattering Transform (ST) and LSTM networks. We used PhysioNet/Computing in Cardiology Challenge 2020 data that includes normal sinus rhythms and eight classes of cardiac arrhythmias.
Methods: The ECG signals were first presented to a translation-invariant ST representation layer, which reduced the sampling rate from fin=500Hz to fout=7.8Hz, that is, in the vicinity of a few samples per typical heart beat. The ST filtering output was passed to a depthwise-separable convolution layer to first mix the filtered electrodes by the scale-sensitive ST paths and then to mix the paths themselves. This was followed by two Long Short-Term Memory (LSTM) layers to capture feature trajectories over time. The final dense layer used binary cross-entropy loss during training to support multiple arrhythmia classes. Predictions were averaged over time and our decision rule chose any class that exceeded the probability threshold p=0.5; otherwise the maximum probability class was chosen. The system was developed with Tensorflow-Keras and the Kymatio ST Keras package.
Results: Cross-validation test results achieved an F2-score, G2-score and geometric mean of 0.5763±0.0329, 0.3568±0.0295 and 0.4626±0.0341, respectively (mean±standard deviation, see Table 1 for per-fold results). Our best hidden test scores were 0.436, 0.270 and 0.343, respectively.
Conclusions: This classifier architecture shows promising results in this first phase of development. In future work, we hope to improve the system with better pipeline normalization for numerical stability; improved ST filtering-padding of batch training to reduce unwanted artifact; use of bidirectional LSTMs; incorporation of the final decision rule into the loss function; and performing a comprehensive hyperparameter search.