Signal Study Finds Most Explainable-AI Methods for ECG Diagnosis Miss Clinically Relevant Signal Regions

Summary

Researchers evaluated 13 post-hoc explainable-AI (XAI) methods used to interpret ECG classification models, testing them against clinically defined regions of diagnostic interest rather than sample-level plausibility alone. Using four binary classifiers trained on the PTB-XL dataset, they compared explanations across low-amplitude segments and high-amplitude QRS morphology patterns. The results show a systematic failure among methods transferred from computer vision: explanations tend to track signal amplitude rather than clinical relevance, with correlations to amplitude reaching Spearman values up to 0.69. For ischemia detection, one common method (LRP-ε) assigned only 4.6% of relevance to the clinically decisive ST segment, compared with 63.8% for an alternative method (LRP-SIGN). Nine of the 13 evaluated methods performed below chance level for at least one cardiac condition, indicating inconsistent reliability. The authors conclude that global, guideline-grounded evaluation can reveal systematic explanation failures invisible in individual sample-level heatmaps.