Arrhythmia Classification of 12-lead Electrocardiograms by Hybrid Scattering-LSTM Networks

Abstract

Electrocardiogram (ECG) analysis is the standard of care for the diagnosis of irregular heartbeat patterns, known as arrhythmias. This paper presents a deep learning system for the automatic detection and multilabel classification of arrhythmias in ECG recordings. Our system composes three differentiable operators: a scattering transform (ST), a depthwise separable convolutional network (DSC), and a bidirectional long short-term memory network (BiLSTM). The originality of our approach is that all three operators are implemented in Python. This is in contrast to previous publications, which precomputed ST coefficients in MATLAB. The implementation of ST on Python was made possible by using a new software library for scattering transform named Kymatio. This paper presents the first successful application of Ky-matio to the analysis of biomedical signals. As part of the PhysioNet/Computing in Cardiology Challenge 2020, we trained our hybrid Scattering-LSTM model to classify 27 cardiac arrhythmias from two databases of 12-lead ECGs: CPSC2018 and PTB-XL, comprising 32k recordings in total. Our team 'Bitscattered' achieved a Challenge metric of 0.536±0.012 over ten folds of cross-validation but this result may be over-optimistic since we were not able to rank and score on the hidden test set.