callmefair: Automating intersectional bias search and mitigation for fair healthcare AI

Algorithmic bias in healthcare machine learning (ML) frequently shows up along overlapping social dimensions—like gender, race, and age—leading to layered disadvantages that approaches looking at just one attribute at a time tend to miss. To tackle this, we present a flexible, model-agnostic framework built specifically to uncover, measure, and reduce intersectional bias in healthcare AI systems. The framework brings together five fairness metrics that have low mutual correlation, combining them into a single composite fairness score (denoted as F) that can capture disparities not only within individual groups but also across combinations of sensitive attributes. Using a multi-dimensional grid search, we systematically test various ML models alongside different bias-mitigation strategies, with explicit attention to how well they perform on intersectional subgroups. To help practitioners make informed choices, we map the trade-offs between fairness and accuracy in visual fairness–accuracy landscapes. We test our approach on three real-world tabular healthcare datasets, uncovering how fairness and predictive performance interact across complex subgroup intersections. Our results highlight a key insight: some bias-mitigation techniques boost fairness for single or pairwise attributes but can actually worsen inequities when more dimensions of identity overlap.