STABLE-BAG: a stability-aware framework for SHAP-based brain age gap decomposition

Brain age modelling estimates chronological age from neuroimaging features, and the brain age gap between predicted and chronological age is used as a marker of atypical brain aging. SHapley Additive exPlanations (SHAP) attributions are widely applied to investigate the regional contributions to individual brain age predictions and their associated gaps, but the attributions depend on the background set against which they are computed, a component that is frequently left unspecified. We show that SHAP attributions exactly decompose the brain age gap only when the explanation baseline coincides with the subject’s chronological age, whereas dataset-level backgrounds generally fail to satisfy this condition because their base value reflects the model’s expected output over the entire reference cohort. We introduce STABLE-BAG, a framework that constructs a subject-specific background from age-matched peers and selects, for each subject, the age window that balances baseline calibration against the stability of the attributions. In a controlled simulation and on the OpenBHB cohort, STABLE-BAG improved the alignment between the explanation baseline and chronological age, reducing the calibration gap relative to fixed and feature-matched backgrounds, producing attributions that remained stable across the cohort while preserving calibration wherever age-matched peers were sufficiently dense. Because the calibration and instability metrics are computed per subject, the framework also flags explanations that are unreliable even when the underlying age prediction is accurate.