New framework makes brain age gap explanations more stable and calibrated

Researchers have introduced a new framework, called STABLE-BAG, that addresses a subtle but consequential problem in brain age modelling: the explanations generated for individual brain age gap predictions can shift depending on how they are computed, and in many analyses that computation is left unspecified. The work, published in the journal Brain Informatics, focuses on making SHAP-based explanations of brain age gaps both mathematically aligned with the quantity being explained and stable across a cohort.

Brain age modelling estimates a person’s chronological age from neuroimaging features. The difference between the predicted age and the actual chronological age, known as the brain age gap, is widely used as a marker of atypical brain aging. To understand which brain regions drive an individual’s gap, researchers commonly apply SHAP, or SHapley Additive exPlanations, attributions, which decompose a model’s prediction into contributions from individual features.

The team behind STABLE-BAG points out that SHAP attributions depend on a background set, the reference data against which attributions are computed, and that this component is frequently left unspecified in applied studies. Their analysis shows that SHAP attributions exactly decompose the brain age gap only when the explanation baseline coincides with the subject’s chronological age. Dataset-level backgrounds generally fail to satisfy this condition, because their base value reflects the model’s expected output over the entire reference cohort rather than the individual’s true age.

To resolve this, STABLE-BAG constructs a subject-specific background drawn from age-matched peers. For each subject, the framework selects an age window that balances two competing requirements: keeping the explanation baseline calibrated to the subject’s chronological age, and maintaining the stability of the resulting attributions. This means the reference population used to explain a prediction is tailored to each individual rather than fixed across the whole dataset.

Compared with fixed backgrounds and feature-matched backgrounds, STABLE-BAG improved the alignment between the explanation baseline and chronological age, reducing the calibration gap. The attributions it produced remained stable across the cohort while preserving calibration wherever age-matched peers were sufficiently dense in the data.

A key design feature is that both the calibration and the instability metrics are computed per subject rather than averaged over the cohort. This allows the framework to flag explanations that are unreliable, even in cases where the underlying age prediction itself is accurate. In other words, a model can predict a person’s age well while still producing an explanation of the brain age gap that should not be trusted, and STABLE-BAG makes that distinction visible.

If the background set is chosen poorly, the regional attribution may not actually decompose the gap it is meant to explain, potentially misleading downstream interpretation.

The article is published open access and was shared early as a peer-reviewed, accepted manuscript carrying a permanent DOI; this version remains subject to further edits before the final Version of Record replaces it. The study’s claims rest on the simulation and the OpenBHB evaluation described by the authors, and the framework’s performance depends on the density of age-matched peers available for each subject in a given cohort.

By making the choice of explanation baseline explicit and adaptive, STABLE-BAG offers researchers a practical route to brain age gap explanations that are both faithful to the predicted quantity and demonstrably stable, while providing a built-in warning system for individual explanations that fall short of either standard.

Subject of Research: A stability-aware framework for computing reliable SHAP-based explanations of brain age gaps from neuroimaging.

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

Article References: Lombardi, A., Fasano, G., Ciuffreda, M., Colafiglio, T., Danese, D., Musicco, F., Di Sciascio, E., & Di Noia, T. (2026). STABLE-BAG: a stability-aware framework for SHAP-based brain age gap decomposition. Brain Informatics. https://doi.org/10.1186/s40708-026-00338-6

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00338-6

Keywords: brain age gap, SHAP, explainability, neuroimaging, machine learning, brain aging, OpenBHB, attribution stability, calibration, Brain Informatics, STABLE-BAG, stability-aware