Source-linked AI summary
ECG arrhythmia classification using a 2-D convolutional neural network
Tae Joon Jun, Hoang Minh Nguyen, Daeyoun Kang, Dohyeun Kim, Daeyoung Kim, Young-Hak Kim
TL;DR
The paper addresses ECG arrhythmia classification with a method that avoids manual signal filtering and feature engineering. It converts ECG beats into grayscale images for a deep two-dimensional CNN, which achieved 99.05% average accuracy and 97.85% average sensitivity on MIT-BIH data.
Problem
Prior ECG arrhythmia methods have limitations including insufficient cross-validation, beat loss during preprocessing, limited arrhythmia coverage, and relatively low classification performance.
Method
The method transforms MIT-BIH ECG beats into grayscale images and classifies them with an optimized two-dimensional CNN using augmentation and cross-validation.
Results
99.05% average accuracy was achieved, alongside 0.989 AUC, 99.57% specificity, 97.85% average sensitivity, and 98.55% positive predictive value.
Takeaways & Limitations
The results indicate that ECG images combined with a CNN can classify eight ECG beat types without manual noise filtering, feature extraction, or feature reduction.
Takeaways & Limitations
The evaluation uses the MIT-BIH arrhythmia database, and random splitting into ten groups can produce unequal test-set sizes because some beats are ignored.
Abstract
from arXiv · showhide
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.
1 Introduction
ECG arrhythmia classification is clinically important, but prior methods have limitations in validation, signal processing, arrhythmia coverage, and practical performance. The paper proposes a two-dimensional CNN using grayscale ECG images to address these issues.
- Cardiovascular diseases are the leading cause of death, making automatic detection of irregular heart rhythms from ECG signals clinically significant.The paper links prolonged arrhythmias with potentially fatal outcomes and emphasizes periodic heart-rhythm monitoring.
- Related work: Prior ECG arrhythmia classifiers commonly used FFNNs, SVMs, RNNs, random forests, or K-NN with manually engineered feature extraction or reduction.Reported approaches include wavelet transforms, Hermite functions, PCA, higher-order statistics, and RR intervals.
- Related work: Existing literature is limited by selective recordings without cross-validation, beat loss during filtering or feature extraction, few arrhythmia types, and relatively low practical performance.
- Proposed approach: The proposed method transforms one-dimensional ECG signals into grayscale images for deep two-dimensional CNN classification, avoiding required noise filtering and feature extraction.Image augmentation enlarges training data, while convolutional and pooling layers learn relevant feature maps.
- Proposed approach: The method uses MIT-BIH recordings, converts each beat into a 128 x 128 grayscale image, and classifies eight ECG beat types.The image may contain neighboring-beat signals, which the CNN is designed to ignore during learning.
2 Methods
The classification pipeline preprocesses MIT-BIH ECG recordings into two-dimensional images and applies a CNN classifier to identify eight ECG beat types.
- The overall pipeline comprises ECG data preprocessing followed by an ECG arrhythmia classifier trained and tested on the MIT-BIH database.The procedure is summarized in Fig. 1.
- ECG signals are transformed into ECG images because the CNN accepts two-dimensional image inputs.
- The CNN classifier performs classification across eight ECG types, including normal beat and seven arrhythmia beats.
2.1 ECG data pre-processing
The preprocessing stage converts ECG beats into standardized 128 x 128 grayscale images centered on the Q-wave peak for eight beat categories.
- Each ECG beat is plotted as an individual 128 x 128 grayscale image centered on its labeled Q-wave peak.The beat is sliced while excluding the first and last 20 ECG signals from the recording segment.
- The preprocessing produces approximately 100,000 images representing eight ECG beat types.Fig. 2 illustrates the normal beat and seven arrhythmia beat categories as grayscale images.
2.2 ECG arrhythmia classifier
The paper uses a two-dimensional CNN that classifies grayscale ECG images, following a VGGNet-like structure and optimizing training to reduce overfitting and improve accuracy. The approach combines image-based augmentation with CNN-specific architectural and training choices.
- ECG arrhythmia classifier: The classifier converts ECG signals into grayscale images so 2-D convolution and pooling can process their spatial locality.The authors relate this representation to the visual process physicians use when judging ECG signals.
- ECG arrhythmia classifier: The proposed model follows VGGNet’s basic structure but is compared with AlexNet and VGGNet rather than deeper GoogLeNet, ResNet, or DenseNet models.The paper describes the ECG images as relatively simple 128 x 128 grayscale images and cites overfitting and additional parameters as reasons for limiting the comparisons.
- Data augmentation: Image-based data augmentation enlarges training data and can maintain class balance, helping address low sensitivity for minority arrhythmia classes.The augmented images are generated inside the model to avoid memory-copy overhead between main memory and GPU memory.
- Cost and optimizer function: Training minimizes a cross-entropy cost with Adam, using a starting learning rate of 0.0001 and exponential decay of 0.95 every 1,000 steps.Validation uses average sensitivity rather than loss, with early stopping after no increase in weighted average sensitivity for 500 global steps.
3 Experiments and Results
The experiments evaluate the proposed CNN on MIT-BIH ECG data using multiple metrics, stratified 10-fold cross-validation, and comparisons with other CNN and prior classification methods. The proposed model achieves strong overall performance, while evaluation design and early-stopping criteria affect interpretation.
- Experimental setup: The evaluation uses MIT-BIH recordings and compares the proposed CNN with AlexNet, VGGNet, and previous ECG arrhythmia classifiers.The database contains approximately 110,000 beats and 15 arrhythmia types; the reported experiment uses selected beat categories.
- Evaluation metrics: Five metrics are considered: AUC, accuracy, specificity, sensitivity, and positive predictive value.AUC is computed from ROC-curve true- and false-positive rates across thresholds.
- Cross-validation: 10-fold stratified cross-validation uses the entire dataset, repeatedly training on nine subsets and testing on the remaining subset.The folds maintain near-even proportions of each data type to reduce variation caused by random train-test splits.
- CNN results: 0.989 AUC, 99.05% average accuracy, 99.57% specificity, 97.85% average sensitivity, and 98.55% positive predictive value are reported for the proposed CNN.With augmentation, the proposed model has the best AUC, accuracy, and sensitivity among the compared CNN models.
- Training behavior: Loss converges near 1,000 steps before the other evaluation metrics, so loss-based early stopping may produce low accuracy and sensitivity under class imbalance.The paper recommends considering AUC or average sensitivity when selecting an early-stopping criterion.
- Comparison with existing approaches: The proposed method reports the best average accuracy, average sensitivity, and average positive predictive value among the compared previous works, with comparisons limited by differing test sets and arrhythmia types.For studies with more than 100,000 test examples, the proposed method reports the best result on every evaluation metric.
4 Discussion and Conclusion
The paper presents two-dimensional CNN classification of ECG images generated from MIT-BIH recordings. It reports strong classification metrics and suggests that the approach could support expert diagnosis and ECG-monitoring systems.
- Method: The proposed method transforms MIT-BIH ECG recordings into 128 x 128 grayscale images for two-dimensional CNN classification.The study obtains over 100,000 ECG beat images spanning eight beat types, including normal beats and seven arrhythmias.
- Results: 0.989 AUC, 99.05% average accuracy, 99.57% specificity, 97.85% average sensitivity, and 98.55% average positive predictive value are achieved.The optimized scheme incorporates data augmentation, regularization, and K-fold cross-validation.
- Implications: The results indicate that ECG images with a CNN can help experts diagnose cardiovascular diseases from ECG signals.The paper also describes possible application to medical robots or scanners that monitor ECG signals.