SHAP Feature Attribution Consistency in Diabetes Prediction: Disentangling Feature Availability from Population and Label Differences

This paper investigates the consistency of SHAP feature attributions for diabetes prediction across clinical (PIMA) and survey-based (BRFSS) datasets. We introduce a category-level consistency analysis to distinguish the effects of feature availability from genuine population differences. The study identifies Age as a directionally consistent and bootstrap-stable predictor across both datasets and demonstrates that apparent disagreement in feature importance is largely driven by differences in measured features rather than underlying population characteristics.