AI-based brain age maps dementia stages

Clinical takeaway: Regional brain age is a sharper research instrument than the single-number version, but its clearest signal shows up in patients already diagnosed with Alzheimer's disease, not in the earlier stages where a marker would change management.

Effective dementia treatment keeps running into a timing problem. Anatomy starts shifting years before a patient reports anything amiss, but by the time cognition slips noticeably, the useful window has narrowed. The concept of brain age is intended to address that gap by reading an MRI for how old tissue looks. Researchers have used AI to make the measure far more anatomically precise, and in the process shown how much clinical precision it still lacks.

Conventional brain-age models return one figure for the whole organ. That average can flag a deviation but cannot locate it, and location carries most of the meaning. Aging turns out to be patterned rather than diffuse: in healthy adults, frontal and temporal regions looked older than the rest of the brain.

Across the cortex, the average brain age gap climbed to 2.9 years in Alzheimer's disease from 1.6 years in mild cognitive impairment and roughly zero in healthy adults. Regionally, the largest gaps separating Alzheimer's patients from healthy adults sat in deep gray matter, about 3.6 years in the pallidum and 3.5 in the putamen, slightly ahead of the right hippocampus and amygdala.

Those margins are small relative to the model's accuracy on an individual, where estimates missed a healthy person's actual age by about six years on average. The group patterns are consistent; a single patient's map is not yet precise enough to read. The link between older-looking regions and cognitive scores held up in the Alzheimer's group across most measures, but faded to non-significance in mild cognitive impairment and in healthy adults after adjusting for sex and education.

A deep learning model was trained on MRIs from nearly 15,000 cognitively healthy adults across six public research datasets, then tested on about 1,900 participants in the Alzheimer's Disease Neuroimaging Initiative, spanning normal cognition, mild cognitive impairment, and Alzheimer's disease.

Participants here were each measured once, so the staged pattern shows how groups differ, not how a brain changes. The question that would make regional brain age clinically useful is whether the maps predict who converts from mild impairment to Alzheimer's disease, and that requires following people over time. The model also needs testing against clinical MRI, which varies far more than research-grade scans.

"By understanding how individual regions age, as well as how those patterns differ from person to person, we're moving toward a much more precise understanding of healthy aging and neurodegenerative disease," said Andrei Irimia, associate professor of gerontology at the USC Leonard Davis School of Gerontology.

Source: Chaudhari NN, et al. (2026 Aug 3) Proc Natl Acad Sci U S A. Deep learning maps local brain aging in relation to cognition across human adulthood