[Paper Review] ECG arrhythmia classification using a 2-D convolutional neural network

Tae Joon Jun, Hoang M. Nguyen+4 more|arXiv (Cornell University)|Apr 18, 2018

ECG Monitoring and Analysis33 references205 citations

TL;DR

The authors transform ECG beats into 128×128 grayscale images and classify eight beat types using a 2-D CNN, achieving high accuracy and validating with 10-fold cross-validation on MIT-BIH data. They compare against AlexNet and VGGNet and employ data augmentation and modern training techniques.

ABSTRACT

In this paper, we propose an effective electrocardiogram (ECG) arrhythmia classification method using a deep two-dimensional convolutional neural network (CNN) which recently shows outstanding performance in the field of pattern recognition. Every ECG beat was transformed into a two-dimensional grayscale image as an input data for the CNN classifier. Optimization of the proposed CNN classifier includes various deep learning techniques such as batch normalization, data augmentation, Xavier initialization, and dropout. In addition, we compared our proposed classifier with two well-known CNN models; AlexNet and VGGNet. ECG recordings from the MIT-BIH arrhythmia database were used for the evaluation of the classifier. As a result, our classifier achieved 99.05% average accuracy with 97.85% average sensitivity. To precisely validate our CNN classifier, 10-fold cross-validation was performed at the evaluation which involves every ECG recording as a test data. Our experimental results have successfully validated that the proposed CNN classifier with the transformed ECG images can achieve excellent classification accuracy without any manual pre-processing of the ECG signals such as noise filtering, feature extraction, and feature reduction.

Motivation & Objective

• Motivate automatic ECG arrhythmia detection to improve diagnostic consistency and enable robust monitoring across devices.
• Propose a 2-D CNN that operates on ECG images to avoid manual feature extraction and noise filtering.
• Evaluate performance against standard CNN architectures and establish strong cross-validation results.

Proposed method

• Transform each ECG beat into a 128×128 grayscale image centered on the Q-wave peak.
• Design and optimize a 2-D CNN inspired by VGGNet with 3×3 kernels and batch normalization, dropout, and Xavier initialization.
• Apply data augmentation by cropping 9 variants per beat to balance classes and improve sensitivity.
• Use Adam optimizer with initial learning rate 0.001 and exponential decay; validate using average sensitivity and stop after stabilization.
• Compare against AlexNet and VGGNet using MIT-BIH data and 10-fold stratified cross-validation.

Experimental results

Research questions

• RQ1 Can transforming 1-D ECG signals into 2-D images enable effective arrhythmia classification with CNNs without manual preprocessing?
• RQ2 Does data augmentation improve sensitivity and overall accuracy for underrepresented arrhythmia classes?
• RQ3 How does the proposed 2-D CNN compare to established CNN architectures (AlexNet, VGGNet) on MIT-BIH data?
• RQ4 What is the impact of training strategies (batch normalization placement, activation choices, initialization) on performance for ECG image inputs?

Key findings

• The proposed 2-D CNN achieves 99.05% average accuracy and 97.85% average sensitivity with 10-fold cross-validation.
• With data augmentation, the model achieves 0.989 AUC, 99.05% accuracy, 97.85% average sensitivity, and 98.55% PPV.
• Compared to AlexNet and VGGNet, the proposed model with augmentation provides the best overall accuracy and AUC, while AlexNet with augmentation shows the best specificity and PPV.
• Sensitivity improvements from augmentation are especially notable for APC, VFW, and VEB classes (gains of 2.89%, 2.75%, and 4.72% respectively).
• The method avoids explicit noise filtering and feature extraction by operating on ECG images, and 10-fold cross-validation ensures robust evaluation across recordings.