Source-linked AI summary
Convolutional Recurrent Neural Networks for Electrocardiogram Classification
Martin Zihlmann, Dmytro Perekrestenko, Michael Tschannen
TL;DR
The paper targets limited generalization in automated AF classification from arbitrary-length single-lead ECG recordings. It compares CNN and CRNN architectures with ECG-specific augmentation, finding that a CRNN ensemble achieved an F1,avg of 0.82 on the private challenge test set, the second-best rounded score.
Problem
Existing AF classification methods have poor generalization because they rely on small or carefully selected data sets.
Method
The paper evaluates a CNN with averaging-based temporal aggregation and a CRNN combining convolutional feature extraction with LSTM temporal aggregation, using ECG data augmentation.
Results
An ensemble of CRNNs achieved an F1,avg of 0.82 on the private challenge testing set, corresponding to the second-best rounded challenge score.
Takeaways & Limitations
The results indicate that LSTM-based temporal aggregation is more effective than averaging for this ECG classification task when data augmentation is used.
Takeaways & Limitations
The CRNN has more parameters and potentially higher model capacity than the CNN, so the aggregation comparison should be interpreted cautiously.
Abstract
from arXiv · showhide
We propose two deep neural network architectures for classification of arbitrary-length electrocardiogram (ECG) recordings and evaluate them on the atrial fibrillation (AF) classification data set provided by the PhysioNet/CinC Challenge 2017. The first architecture is a deep convolutional neural network (CNN) with averaging-based feature aggregation across time. The second architecture combines convolutional layers for feature extraction with long-short term memory (LSTM) layers for temporal aggregation of features. As a key ingredient of our training procedure we introduce a simple data augmentation scheme for ECG data and demonstrate its effectiveness in the AF classification task at hand. The second architecture was found to outperform the first one, obtaining an $F_1$ score of $82.1$% on the hidden challenge testing set.
1. Introduction
The paper addresses AF classification from single-lead ECG recordings, motivated by the clinical burden of AF and limited generalization of existing methods. It proposes CNN and CRNN architectures with ECG-specific data augmentation for arbitrary-length recordings.
- Existing AF classification methods have poor generalization because they are trained or evaluated on small or carefully selected data sets.
- The paper proposes two deep neural network architectures for classifying arbitrary-length ECG recordings on the PhysioNet/CinC Challenge 2017 AF data set.
- The CNN uses averaging-based temporal feature aggregation, whereas the CRNN combines convolutional feature extraction with a three-layer LSTM for temporal aggregation.
- The training procedure introduces a simple data augmentation scheme tailored to the ECG data.
- The architectures are loosely inspired by prior CRNN and LSTM applications in sound detection, EEG classification, electronic health records, and ECG arrhythmia detection.
2. Methods1
The method preprocesses arbitrary-length ECG signals as logarithmic spectrograms, extracts features with convolutional blocks, aggregates them across time, and classifies the resulting representations. CNN and CRNN variants differ mainly in temporal aggregation, while training uses augmentation and cross-validation.
- 2.1. Network architectures: Both architectures compute a logarithmic spectrogram, extract features with convolutional layers, aggregate across time, and apply a linear classifier.
- 2.1. Network architectures: Preliminary experiments found that the logarithmic transform considerably increased classification accuracy; spectrograms used 64-sample Tukey windows with 50% overlap.
- 2.1. Network architectures: The CNN aggregates variable-length convolutional outputs by averaging, whereas the CRNN flattens feature maps and uses a three-layer bidirectional LSTM with 200 neurons per layer.
- 2.1. Network architectures: Averaging smooths features but may miss short episodic phenomena, while nonlinear LSTM aggregation may preserve them better.
- 2.2. Training procedures: The CRNN required a three-phase training protocol because joint training from scratch did not converge.
- 2.2. Training procedures: Data augmentation was introduced after severe overfitting was observed, with dropout bursts modeling weak signals and random resampling varying simulated heart rates between 60 and 120 bpm.
- 2.3. Evaluation: Evaluation used 8,528 single-lead ECG recordings, stratified five-fold cross-validation, and an average of F1 scores for normal, AF, and other rhythms.
3. Results
Data augmentation substantially improved overall F1 for both architectures, with a slightly larger improvement for the CRNN. The selected CRNN ensemble achieved an F1,avg of 0.82 on the private challenge test set, the second-best rounded score.
- Data augmentation considerably increased F1,avg for both CNN and CRNN, with a slightly better improvement for the CRNN.
- With augmentation, the CRNN achieved slightly higher F1,avg and overall accuracy than the CNN, whereas without augmentation the CNN had the advantage.
- An ensemble of CRNNs obtained an F1,avg of 0.82 on the private challenge testing set.
- The ensemble’s score was the second best result after rounding to two decimal places, while its average computation use was 58.1% of the available quota.
4. Discussion
The results favor LSTM-based temporal aggregation over averaging when data augmentation is used, but this comparison is confounded by the CRNN’s greater parameter count. Additional training refinements and validation-set strategies did not consistently improve performance.
- LSTM-based feature aggregation was more effective than averaging for ECG classification when data augmentation was employed.The authors caution that the CRNN has more parameters and potentially greater model capacity than the CNN.
- The CRNN’s advantage over the CNN may partly reflect its higher model capacity because it contains more parameters.
- Continuing training on the combined training and validation sets decreased F1,avg in challenge submissions.This production-model strategy was explored as a computationally and storage-efficient alternative to ensembling.
- Pretraining convolutional components on the PTB Diagnostic ECG Database did not improve results compared with random-weight initialization.The database contained 549 14-lead ECG recordings from 290 subjects with varied cardiac conditions.
5. Conclusion
The paper reports successful development and evaluation of two deep neural network architectures and an effective ECG data augmentation scheme. Future work includes extending the architectures to multi-lead ECG and refining augmentation, including heart-rate-aware resampling.
- The study developed and evaluated two deep neural network architectures for ECG classification.
- The proposed data augmentation scheme was demonstrated to be effective for ECG data.
- Applying the architectures to multi-lead ECG data and different pathology is identified as a future research direction.
- Refining augmentation could incorporate actual heart rate into random resampling instead of assuming 80 bpm.