Racial and Ethnic Disparities in Brain Age Algorithm Performance: Investigating Bias Across Six Popular Methods

Brain age algorithms, which estimate biological aging from neuroimaging data, are increasingly used as biomarkers for health and disease. However, most algorithms are trained on datasets with limited racial and ethnic diversity, raising concerns about potential algorithmic bias that could exacerbate health disparities. To probe this potential, we evaluated six popular brain age algorithms using data from the Health and Aging Brain Study-Health Disparities (HABS-HD), comprising 1,123 White American, 1,107 Hispanic American, and 678 African American participants, ages [≥]50. Comparing correlations between brain age and chronological age across racial/ethnic groups, relations were consistently weaker for African American participants compared to White and Hispanic American participants across most algorithms (ranging from r=0.51-0.85 for African Americans vs. r=0.57-0.89 for other groups). We also examined error for brain age v. chronological age and found significant differences in median errors across racial/ethnic groups, though specific patterns varied by algorithm. Sensitivity models weighting for age, sex, and scan quality noted similar patterns, with all algorithms maintaining significant differences in correlation or median prediction error between groups. Our findings reveal systematic performance differences in brain age algorithms across racial and ethnic groups, with most algorithms consistently showing reduced algorithm accuracy for African American and/or Hispanic-American participants. These biases, which are likely introduced at multiple stages of algorithm development, could impact clinical utility and diagnostic accuracy. Results highlight the urgent need for more inclusive algorithm development and validation to ensure equitable healthcare applications of neuroimaging biomarkers.