FRAME: separating sampling variation from representational cause in medical imaging fairness

Medical imaging models often exhibit performance disparities across demographic groups, a phenomenon typically attributed to the model learning and encoding sensitive attributes like race or age. Standard fairness interventions aim to mitigate these gaps by erasing demographic information from the model's internal representations. This paper introduces FRAME (Fair-model Reference And Mechanism Evaluation), a two-step framework designed to audit these claims. First, FRAME calculates a fair-model reference—the performance difference expected due to sampling variation alone given the observed subgroup sizes. Second, it uses representation-space interventions to test whether the remaining performance gap is causally linked to demographic encoding.

Across 702,206 images and 36 encoders, the fair-model reference accounts for a median 41% of reported race differences and 22% of age differences. For many combinations of model and finding, the reported difference does not exceed what would be expected under exact fairness. When a gap does exceed this reference, injecting demographic decodability into the model's features does not change the performance difference. Furthermore, no standard fairness intervention tested—including adversarial removal, distributionally robust optimization, and concept erasure—reduced the performance gap more than a simple change of the pretraining random seed. While image-text pretraining improved worst-group performance, these gains were independent of demographic erasure.

These findings challenge the prevailing belief that demographic encoding is the primary driver of subgroup performance disparities in medical imaging. By quantifying the role of sampling variation, FRAME reveals that many reported "biases" are statistically expected outcomes of small subgroup sizes rather than evidence of model-level discrimination. This suggests that the field should prioritize larger, more representative cohorts and mechanistic explanations over interventions that remove demographic features, which often fail to close the performance gap while potentially degrading overall model utility.