What does this research mean for the field?

Attribution methods for ECG classification demonstrate limited reliability, high variability, and incomplete dependence on learned parameters, constraining their clinical utility in interpreting deep neural network predictions. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

The research aim is to assess the reliability of various attribution methods used for explaining models in ECG classification.

March 13, 2026Open Access

Signal or noise? Evaluating commonly used attribution methods for explaining deep neural networks in electrocardiogram classification

Key Result

Attribution methods for ECG classification demonstrated limited reliability and high variability, constraining their clinical utility in interpreting deep neural network predictions.

Key Points

The research aim is to assess the reliability of various attribution methods used for explaining models in ECG classification.
Analyzed 12 attribution methods using a dataset of 873,710 ECGs across nine diagnostic classes.
Applied methods to convolutional neural network-based models trained for ECG classification.
Conducted four performance evaluations: inter-method similarity, self-consistency, dependence on model weights, and feature identification.
Task models achieved an area under the receiver operating curve above 0.95.
Attribution methods showed low correlation and high variability across comparisons.
Self-consistency for most methods was moderate (mean correlation 0.41–0.65).
Randomizing model weights resulted in significant loss of correlation, with some methods retaining some stability.

Structured PICO

Population

873,710 median beat ECGs spanning nine diagnostic classes

Intervention

12 attribution methods applied to convolutional neural network-based models trained for ECG classification

Outcome

Performance evaluated across four experiments: inter-method similarity, self-consistency, dependence on model weights, and ability to identify features important for model inference

Attribution methods for explaining deep neural networks in ECG classification demonstrate limited reliability and instability, constraining their utility in clinical healthcare settings.

Main Result

Absolute Event Rate: 0% vs 0%

Abstract

Abstract Background and Aims Attribution-based explainability methods are widely used in electrocardiogram (ECG) analysis to interpret predictions from “black-box” deep neural networks (DNNs). To be useful in clinical applications, attribution methods must produce explanations that are both clear and reflective of the model’s inner workings. This study evaluates 12 attribution methods in DNN-based ECG classification. Methods We analysed 12 attribution methods using a dataset of 873,710 median beat ECGs spanning nine diagnostic classes. Methods were applied to convolutional neural network-based models trained for ECG classification. Performance was evaluated across four experiments: inter-method similarity, self-consistency, dependence on model weights, and ability to identify features important for model inference. Results All task models achieved an area under the receiver operating curve above 0.95. Attribution methods demonstrated low correlation and high variability across inter-method comparisons. Self-consistency across random model initialisations was moderate for most methods (mean correlation 0.41–0.65). Randomising model weights led to rapid loss of correlation, although some methods did not converge to zero. Perturbation of input data revealed differences in how well attribution methods identified features relevant to model performance. Conclusions Attribution methods demonstrated limited reliability, instability across model variants and incomplete dependence on learned parameters, constraining their utility in high-stakes settings such as healthcare. These findings suggest that attribution techniques should be used cautiously and supported by task-specific sanity checks. Approaches grounded in rigorous validation, inherently interpretable modelling or counterfactual explanations may better support clinically meaningful insight.