Spectral Entropy Quantifies XAI Noise in ECG Data.

Explainable AI (XAI) techniques are increasingly vital for understanding and trusting deep learning models, especially in sensitive domains like healthcare where decisions must be justified. However, XAI tools often employ heuristics that can introduce extraneous signal noise into the explanations, making it difficult to discern what truly originates from the model versus what is an artifact of the XAI method itself. This research introduces a novel approach to address this challenge by proposing the use of spectral entropy as a quantitative measure of noise within XAI outputs. The study demonstrates the effectiveness of this method in the context of classifying cardiac arrhythmias using Electrocardiogram (ECG) data. By applying spectral entropy to the outputs of various post-hoc explainability techniques, the researchers show how it can help identify and quantify the noise component, thereby improving the clarity and reliability of model explanations.