[Paper Review] A Comparison of 1-D and 2-D Deep Convolutional Neural Networks in ECG Classification

Yunan Wu, Yang Feng+3 more|arXiv (Cornell University)|Oct 16, 2018

ECG Monitoring and Analysis8 references73 citations

TL;DR

The paper compares 1-D ECG signal input with 2-D ECG image input using AlexNet-like CNNs, showing that 2-D CNNs initialized with ImageNet weights achieve 98% accuracy on MIT-BIH and remain robust across 20–35 dB SNR, especially when data are limited.

ABSTRACT

Effective detection of arrhythmia is an important task in the remote monitoring of electrocardiogram (ECG). The traditional ECG recognition depends on the judgment of the clinicians' experience, but the results suffer from the probability of human error due to the fatigue. To solve this problem, an ECG signal classification method based on the images is presented to classify ECG signals into normal and abnormal beats by using two-dimensional convolutional neural networks (2D-CNNs). First, we compare the accuracy and robustness between one-dimensional ECG signal input method and two-dimensional image input method in AlexNet network. Then, in order to alleviate the overfitting problem in two-dimensional network, we initialize AlexNet-like network with weights trained on ImageNet, to fit the training ECG images and fine-tune the model, and to further improve the accuracy and robustness of ECG classification. The performance evaluated on the MIT-BIH arrhythmia database demonstrates that the proposed method can achieve the accuracy of 98% and maintain high accuracy within SNR range from 20 dB to 35 dB. The experiment shows that the 2D-CNNs initialized with AlexNet weights performs better than one-dimensional signal method without a large-scale dataset.

Motivation & Objective

• Motivate reliable automatic ECG arrhythmia detection to reduce human error due to fatigue.
• Investigate whether 2-D image representations with CNNs outperform 1-D signal inputs in ECG classification.
• Evaluate the impact of ImageNet pretraining on 2-D ECG classifiers to mitigate overfitting with limited data.

Proposed method

• Convert ECG signals to image representations for 2-D CNN input.
• Compare accuracy and robustness between 1-D ECG signal input and 2-D image input using an AlexNet-like network.
• Initialize the 2-D network with weights pretrained on ImageNet and fine-tune on ECG images.
• Evaluate performance on the MIT-BIH arrhythmia database.
• Assess robustness across a signal-to-noise ratio (SNR) range from 20 dB to 35 dB.

Experimental results

Research questions

• RQ1 Does a 2-D image-based CNN outperform a 1-D signal-based CNN for ECG classification under limited training data?
• RQ2 Does ImageNet pretraining help 2-D ECG classifiers achieve higher accuracy and robustness?
• RQ3 How does performance vary with noise levels (SNR 20–35 dB) for the two input representations?

Key findings

• 2-D CNNs initialized with ImageNet weights achieve higher accuracy than 1-D input methods.
• On MIT-BIH, the proposed 2-D method reaches 98% accuracy.
• The 2-D approach maintains high accuracy within the SNR range of 20–35 dB.
• ImageNet-pretrained 2-D networks help alleviate overfitting when training data are not large.
• The study demonstrates benefits of image-based ECG classification when large labeled datasets are unavailable.