Visual interpretation of deep learning model in ECG classification: A comprehensive evaluation of feature attribution...

Area of Science:

• Artificial Intelligence
• Cardiology
• Medical Imaging Analysis

Background:

• Feature attribution methods are increasingly used in deep learning for electrocardiogram (ECG) classification to aid clinical understanding and model reliability.
• A lack of comprehensive studies has led to the selection of feature attribution methods for ECG data without a clear understanding of their suitability.

Purpose of the Study:

• To conduct a large-scale assessment of eleven popular feature attribution methods for ECG classification.
• To identify the most effective feature attribution methods for analyzing deep learning models in the context of ECG data.

Main Methods:

• Evaluated eleven feature attribution methods on five large ECG datasets using a ResNet-18 model.
• Performed both automatic evaluations using annotated datasets and human evaluations with three cardiac experts.

Main Results:

• Guided Grad-CAM, especially using its absolute values, demonstrated the best performance among the evaluated methods.
• Cardiac experts validated that Guided Grad-CAM can identify diagnostically relevant electrophysiological characteristics in ECGs.

Conclusions:

• Guided Grad-CAM is a highly effective feature attribution method for ECG classification tasks.
• While effective, the performance of Guided Grad-CAM varied across the 17 different diagnoses investigated, suggesting further research is needed for specific conditions.