Ho, Ngoc-Huynh; Charisis, Sokratis; Honnorat, Nicolas; Brandigampala, Sachintha Ransara; Wang, Di; Heckbert, Susan R.; Fox, Peter T.; Martinez, David; Wang, David H.; Hughes, Timothy M.; Archer, Derek B.; Hohman, Timothy J.; Seshadri, Sudha; Davatzikos, Christos; Habes, Mohamad. (2026). . Nature Communications, 17(1), 8026.
Dementia affects millions of people worldwide and is projected to triple by 2050, making early and accurate diagnosis important for treatment and quality of life. However, artificial intelligence models used to classify dementia from MRI scans may not perform equally well across racial and ethnic groups. This study examined differences in dementia classification among 6,584 Non-Hispanic White, 1,263 Non-Hispanic African American, and 713 Hispanic White individuals. The researchers found significant differences in model performance, particularly when a model trained using data from one group was applied to another. To reduce these differences, they evaluated RegAlign, a machine-learning approach that uses a small amount of data from underrepresented groups to help the model adapt while accounting for differences between groups and dementia categories. RegAlign substantially reduced performance gaps, especially between Non-Hispanic White and Hispanic populations. These findings highlight the importance of using diverse training data and fairness-focused machine-learning methods to improve the accuracy and consistency of MRI-based dementia classification across different racial and ethnic populations.
Fig. 1: Baseline performance on group-balanced subsets for assessing disparities across multi-racial and multi-ethnic populations.