Introducing δ-XAI: a novel sensitivity-based method for local AI explanations

A new tool called delta-XAI helps explain AI decisions in healthcare: The method provides clear, human-understandable explanations for why AI models make specific predictions for individual patients, addressing a major barrier to clinicians trusting AI systems in medical settings.

The technique is particularly good at identifying which patient factors matter most: Delta-XAI shows higher sensitivity than existing methods (like Shapley values) in detecting which features have the strongest impact on predictions, especially when dealing with extreme or unusual patient values.

It works for both prediction types doctors encounter: The delta-XAI method can explain AI decisions in both regression problems (predicting continuous values like blood pressure) and classification problems (predicting disease presence/absence), making it broadly applicable across clinical scenarios.

Performance is generally reliable but has important limitations: While delta-XAI results are generally consistent with established benchmark methods, the researchers found notable differences when models encounter extreme feature values, suggesting clinicians should be cautious when applying explanations to unusual cases.

Responsible AI adoption requires practical, implementable solutions: By providing code for implementation alongside the method, the researchers are supporting the real-world deployment of explainable AI—a critical step toward building the transparency and trust needed for AI integration in clinical practice.