Researchers from the South Texas Alzheimer’s Disease Research Center (ADRC) published a study examining fairness and accuracy in MRI-based dementia classification across diverse populations. Using data from 8,560 participants, including Non-Hispanic White, Non-Hispanic African American, and Hispanic White individuals, the team identified significant performance disparities when machine learning models trained on one racial or ethnic group were applied to another.
To address these biases, the researchers evaluated RegAlign, an advanced domain adaptation framework designed to improve model performance in underrepresented populations. The approach combines focal learning, class-weighted supervision, and class-conditional alignment to enhance cross-group generalizability. Results demonstrated a substantial reduction in classification performance gaps, particularly between Non-Hispanic White and Hispanic populations.
The findings underscore the importance of fairness-aware artificial intelligence methods and the inclusion of diverse populations in training datasets to improve the accuracy, equity, and clinical reliability of MRI-based dementia diagnosis as the global burden of dementia continues to grow.