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

Researchers propose using spectral entropy to measure noise introduced by explainability AI (XAI) techniques applied to deep learning models, demonstrating the approach on ECG arrhythmia classification. The work addresses a critical gap in healthcare AI where distinguishing between genuine model signals and XAI-generated artifacts is essential for clinical trust and safety.

This research tackles a fundamental challenge in medical AI deployment: validating the trustworthiness of explanations generated by black-box models. Healthcare systems increasingly rely on deep learning for diagnostic tasks like arrhythmia detection, yet clinicians must understand model reasoning before adopting these tools. Current explainability methods—such as attention maps or gradient-based visualizations—often introduce artifacts that obscure the actual decision-making process, creating false confidence in flawed explanations.

The authors' application of spectral entropy as a noise quantification metric represents a practical advancement in XAI validation. Spectral entropy measures signal disorder in frequency space, providing an objective framework to distinguish meaningful interpretations from XAI-generated distortions in ECG data. This approach is particularly relevant because ECG signals have well-defined spectral characteristics, making noise contamination measurable and comparable across different XAI techniques.

The implications extend beyond ECG analysis. As regulatory bodies like the FDA and EU's AI Act demand explainability for high-stakes medical decisions, organizations need rigorous methods to validate XAI outputs. This research provides a quantitative foundation that could inform standards for acceptable explanation quality in clinical contexts. Healthcare institutions deploying AI systems can use such metrics to benchmark different explainability tools and select those introducing minimal noise.

Future work should explore whether spectral entropy generalizes to other medical signals (EEG, EMG) and non-medical domains. Standardizing XAI noise measurement could accelerate clinical AI adoption by providing objective criteria for model trustworthiness, directly supporting regulatory approval processes and clinical integration timelines.

• →Spectral entropy provides a quantitative measure for assessing noise introduced by explainability techniques in AI models.
• →Healthcare AI systems require validated explainability to ensure clinical trust and regulatory compliance.
• →XAI tools often add artifacts that obscure genuine model signals, making noise measurement critical for model validation.
• →The method demonstrates measurable applicability in ECG arrhythmia classification across different post hoc explanation techniques.
• →Standardized XAI noise metrics could accelerate FDA approval and clinical adoption of deep learning diagnostic tools.