Analyzing the impact of transfer learning on explanation accuracy in deep learning-based ECG recognition systems

Abstract

Electrocardiogram (ECG) recognition systems now play a leading role in the early detection of cardiovascular diseases. However, the explanation of judgments made by deep learning models in these systems is prominent for clinical acceptance. This article reveals the effect of transfer learning in ECG recognition systems on decision precision. This article investigated the role of transfer learning in ECG image classification using a customized convolutional neural network (CNN) with and without a VGG16 architecture. The customized CNN model with the VGG16 achieved a good test accuracy of 98.40%. Gradient-weighted class activation mapping (Grad-CAM), for this model, gave the wrong information because it focused on parts of the ECG that were not important for making decisions instead of features necessary for clinical diagnosis, like the P wave, QRS complex, and T wave. A proposed model that only used customized CNN layers and did not use transfer learning performed 99.08% on tests gave correct Grad-CAM explanations and correctly identified the influencing areas of decision-making in the ECG image. Because of these results, it seems that transfer learning might provide good performance metrics, but it might also make things harder to understand, which could make it harder for deep learning models that use ECG recognition to be reliable for diagnosis. This article concludes with a call for careful consideration when using transfer learning in the medical field, as model explanations resulting from such learning may not be appropriate when it comes to domain-specific interpretations.

Keywords

• convolutional neural networks
• gradient-weighted class activation mapping
• vgg16
• electrocardiogram
• transfer learning