When Brain Models Are not Universal: Benchmarking of Ethnic Bias in MRI-Based Cognitive Prediction Across Modalities

Abstract

Predictive neuroimaging models promise precision medicine but risk exacerbating health inequities if they perform unevenly across ethnic/racial groups. Using the Adolescent Brain Cognitive Development data, we benchmarked ethnic/racial bias in models predicting cognitive functioning from 91 MRI phenotypes across four training strategies. Models trained on one ethnicity performed best within that group. Models trained on participants sampled without regard to ethnicity, a common practice, performed better on White participants, likely because the ABCD sample was predominantly White. Training on equal-sized White and African American subsamples reduced disparities without accuracy loss, emerging as the upper bound for both accuracy and fairness. Structural MRI exhibited the greatest bias, whereas task-based fMRI phenotypes were more equitable. Stronger brain-cognition associations generalized more equitably, but multimodal stacking--despite enhancing prediction--did not improve fairness. Increasing representation of African American participants improved performance up to balanced sampling, with diminishing returns beyond. This first modality-wide benchmark reveals pervasive, modality-dependent ethnic bias in cognitive prediction and identifies key factors shaping equity in neuroimaging models.