Evaluating gradient-based explanation methods for neural network ECG analysis using heatmaps.

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Signal Processing

Background:

• Deep neural networks (DNNs) are increasingly used for electrocardiogram (ECG) analysis.
• Interpreting DNN predictions is crucial for clinical adoption and trust.
• Heatmap visualizations are common for explaining DNN outputs.

Purpose of the Study:

• To evaluate popular explanation methods for DNN-based ECG analysis.
• To compare qualitative expert assessments with objective perturbation-based evaluations.
• To provide recommendations for selecting explanation methods in medical AI.

Main Methods:

• A residual DNN was trained for ECG interval and amplitude prediction.
• Nine explanation methods (Saliency, Deconvolution, Guided backpropagation, Gradient SHAP, SmoothGrad, Input × gradient, DeepLIFT, Integrated gradients, GradCAM) were assessed.
• Methods were evaluated qualitatively by medical experts and quantitatively via perturbation.

Main Results:

• No single explanation method consistently outperformed others; some were inferior.
• Significant disagreement was observed between human expert evaluation and objective perturbation-based evaluation.
• Method performance varied depending on the specific ECG measure being predicted.

Conclusions:

• The optimal explanation method is context-dependent on the ECG measure.
• Collaboration between data scientists and medical experts is vital for developing useful explanation methods.
• Employing multiple explanation methods is advised to ensure robust and reliable interpretations in medical AI.