Quantifying Explainable AI-introduced signal noise on ECG data with Spectral Entropy

Explainability tools can introduce their own distortions: The paper reveals that techniques used to explain AI decisions in healthcare—meant to build trust—can actually add "noise" or artifacts to the explanations themselves, making it harder to distinguish what the model actually learned from what the explanation method added.

Current XAI methods lack quality metrics: There is no standard way to measure whether an explanation is reliable or contaminated by the explanation technique itself. The authors propose using "spectral entropy" as a measurable standard for assessing explanation quality in medical AI systems like ECG analysis.

The problem is particularly acute in healthcare: Because medical AI decisions directly affect patient care and require clinical justification, having noisy or unreliable explanations undermines the entire purpose of using explainability tools—which is to enable doctors to trust and validate AI recommendations.

Different explanation techniques perform differently: The research demonstrates that various post hoc explainability methods (techniques applied after a model makes a decision) introduce varying amounts of noise when explaining arrhythmia classifications, suggesting some methods are more trustworthy than others.

Responsible AI requires measuring explanation reliability: Before deploying explainable AI in clinical settings, organizations should validate that their explanation methods are actually providing clear, noise-free insights rather than potentially misleading doctors with artifacts introduced by the explanation technique itself.