RegAlign cuts by about half the dementia cases that AI does not find

A team with Mohamed Habes and Ngoc-Huynh Ho as leaders did a test of AI that finds dementia in brain scans from 9,007 participants. The AI was less correct for non-Hispanic African American and Hispanic patients than for white patients. A new formula makes the difference less, and the team showed the study on 6 October 2026.

How it works

The team is at the UT San Antonio Health Science Center. The AI measures the volume of brain regions in each scan. The formula, RegAlign, gives more weight to the data of minority patients and moves the predictions to look the same across groups.

The numbers

• Cases that the AI did not find fell from 30.00% to 14.72% for African American patients.
• They fell from 25.88% to 11.13% for Hispanic patients.

The tests used external datasets. A balanced set of scans gave the same difference.

What is not known

After RegAlign, the model looks at different brain parts for each group. Habes thinks that vascular burden, which is damage to blood vessels in the brain, can change how dementia shows on a scan. The team works on more groups and on less usual types of dementia.

Insight: One score hides where medical AI is not correct

This model had a difference that only a view by group could show. In a test of 100 medical tasks, the top agent completed 46% in one try, but only 28% in all 3 tests. These 2 results show that one score is not sufficient, and the buyer that wants results by group and from many tests sets the test. Put that into the contract.