Towards a Shapley Value Graph Framework for Medical peer-influence

Feature importance alone is insufficient for responsible intervention: The paper identifies a critical gap in current explainable AI methods—they show which features matter for predictions but don't explain what happens when you intervene on those features, which is essential for making safe decisions in medical settings.

Graph-based explanations can reveal unintended consequences: By mapping feature-to-feature interactions through a "peer influence framework," the approach aims to show how changing one variable might cascade through a system, helping practitioners anticipate secondary effects of interventions that simple feature importance scores would miss.

Medical AI requires deeper interpretability: The focus on medical peer-influence suggests that healthcare applications demand more sophisticated explanations than general-purpose XAI methods currently provide, since clinical interventions based on incomplete understanding of feature relationships could harm patients.

Black-box models need relational context, not just rankings: Rather than just ranking features by importance, the framework proposes representing how features influence each other, offering a more complete picture of model logic that could better support clinicians in understanding and trusting AI recommendations.