Assessing the Reliability of Machine Learning Explanations in ECG Analysis Through Feature Attribution

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Machine Learning Interpretability

Background:

• Feature attribution methods are widely used to explain decisions of convolutional neural networks (CNNs).
• These methods primarily offer local explanations, hindering systematic evaluation of their global meaningfulness.
• This limitation can foster confirmation bias in model interpretation, particularly in critical applications like healthcare.

Purpose of the Study:

• To systematically investigate the global meaningfulness of feature attribution methods in electrocardiogram (ECG) time series analysis.
• To evaluate the performance of various feature attribution techniques on R-peak, T-wave, and P-wave components of ECG signals.
• To assess these methods across different CNN architectures and explainability frameworks.

Main Methods:

• Utilized a simulated ECG dataset with controlled modifications to R-peak and T-wave features.
• Applied multiple feature attribution techniques to two distinct CNN architectures.
• Extended the evaluation to real-world ECG data to assess practical performance and clarity.

Main Results:

• Feature attribution maps successfully identified diagnostically significant regions in ECG signals.
• Despite highlighting relevant areas, the generated attribution maps exhibited a lack of clarity and produced blurry representations.
• This lack of clarity persisted even under simulated ideal conditions, indicating inherent limitations of current methods.

Conclusions:

• Current feature attribution methods, while capable of localizing important ECG signal components, provide insufficient global interpretability.
• The inherent blurriness in feature attribution maps limits their practical utility for precise clinical decision support.
• Further research is needed to develop more robust and interpretable feature attribution techniques for time-series data like ECGs.