Deep residual 2D convolutional neural network for cardiovascular disease classification

Abstract

Cardiovascular disease (CVD) continues to be a major global health concern, underscoring the need for advancements in medical care. The use of electrocardiograms (ECGs) is crucial for diagnosing cardiac conditions. However, the reliance on professional expertise for manual ECG interpretation poses challenges for expanding accessible healthcare, particularly in community hospitals. To address this, there is a growing interest in leveraging automated and AI-driven ECG analysis systems, which can enhance diagnostic accuracy and efficiency, making quality cardiac care more accessible to a broader population. In this study, we implemented a novel deep two-dimensional convolutional neural network (2D-CNN) on a dataset of PTB-XL for cardiac disorder detection. The studies were performed on 2, 5, and 23 classes of cardiovascular diseases. The our network in classifying healthy/sick patients achived an AUC of 95% and an average accuracy of 87.85%. In 5-classes classification, our model achieved an AUC of 93.46% with an average accuracy of 89.87%. In a more complex scenario involving classification into 23 different classes, the model achieved an AUC of 92.18% and an accuracy of 96.88%. According to the experimental results, our model obtained the best classification result compared to the other methods based on the same public dataset. This indicates that our method can aid healthcare professionals in the clinical analysis of ECGs, offering valuable assistance in diagnosing CVD and contributing to the advancement of computer-aided diagnosis technology.

Figure 1

Distribution of superclasses in the PTB-XL dataset.

Figure 2

Distribution of subclasses in the PTB-XL dataset.

Figure 3

Examples of rhythm ECG signals using lead II.

Figure 4

The proposed deep learning network architecture for automatic classification of cardiovascular diseases.

Figure 5

The process of extracting features in both temporal and spatial analyses from a signal.

Figure 6

Confusion matrices for our method on test dataset for 2 classes.

Figure 7

Confusion matrices for our method on test dataset for 5 classes.

Figure 8

The first part of the confusion matrices depicting the performance of our method on the test dataset for 23 classes.

Figure 9

The second part of the confusion matrices depicting the performance of our method on the test dataset for 23 classes.

Figure 10

ROC curves for 2, 5, and 23 classes.