Robustness of convolutional neural networks to physiological electrocardiogram noise

Abstract

The electrocardiogram (ECG) is a widespread diagnostic tool in healthcare and supports the diagnosis of cardiovascular disorders. Deep learning methods are a successful and popular technique to detect indications of disorders from an ECG signal. However, there are open questions around the robustness of these methods to various factors, including physiological ECG noise. In this study, we generate clean and noisy versions of an ECG dataset before applying symmetric projection attractor reconstruction (SPAR) and scalogram image transformations. A convolutional neural network is used to classify these image transforms. For the clean ECG dataset, F1 scores for SPAR attractor and scalogram transforms were 0.70 and 0.79, respectively. Scores decreased by less than 0.05 for the noisy ECG datasets. Notably, when the network trained on clean data was used to classify the noisy datasets, performance decreases of up to 0.18 in F1 scores were seen. However, when the network trained on the noisy data was used to classify the clean dataset, the decrease was less than 0.05. We conclude that physiological ECG noise impacts classification using deep learning methods and careful consideration should be given to the inclusion of noisy ECG signals in the training data when developing supervised networks for ECG classification. This article is part of the theme issue ‘Advanced computation in cardiovascular physiology: new challenges and opportunities’.

Detecting paroxysmal atrial fibrillation from normal sinus rhythm in equine athletes using Symmetric Projection Attractor Reconstruction and machine learning

“…Subtle differences could be visually discerned between PAF and control cases in the SPAR attractor image, and we elected to use machine learning on metrics drawn from the image to emphasize these. While the visual output of SPAR may indicate the use of deep learning on the attractor images, as Aston and colleagues 34 and Venton and colleagues 35 have demonstrated, we did not have enough data for such an approach, and therefore applied alternative machine learning algorithms using features extracted from the attractors. Furthermore, the 2 classes (PAF and control) were inherently imbalanced owing to the low prevalence of PAF, which we addressed with a synthetic oversampling of the PAF data, since machine learning on imbalanced classes can lead to poor performance.…”

Investigating the Robustness of Deep Learning to Electrocardiogram Noise

“…This dataset is hereafter referred to as the raw dataset. For full details of the data used, see [7].…”

“…When noisy or perturbed input data is passed to a trained ECG model, this can affect performance of the model. Existing studies have looked at ECG perturbations due to adversarial attack [6] and physiological noise [7]. Here, we explore the robustness of trained models to physiological noise further, looking at different levels of noise, and whether different types of physiological ECG noise (ambulatory, non-ambulatory) have a different impact.…”

A Robustness Evaluation of Machine Learning Algorithms for ECG Myocardial Infarction Detection

“…According to our research, there have been few or almost no studies in recent years that evaluated ECG signal classification model robustness. In this study [ 42 ], the authors employed a deep learning approach—a convolutional neural network—to distinguish three classes (normal, AF, and STD) by applying transformed ECGs to images (attractor reconstruction, etc.). They used a random SNR between 5 and 10 dB to assess the durability of their classifier trained on clean data against three forms of noises (em, bw, and ma).…”