Brain age adds something in subjective decline, and nothing in MCI

Design

Retrospective cohort study with Cox models adjusted for age, sex, MMSE and visual rating scales, plus interaction testing

Population

799 memory clinic patients (412 subjective cognitive decline, 387 mild cognitive impairment), mean age 63.3 years, median follow-up 3.2 years

Primary outcome

Progression to dementia

Effect

235 (29.4%) progressed. Brain-PAD hazard ratio 1.04 (95% CI 1.02-1.06) overall; 1.09 (1.03-1.15) in subjective cognitive decline and 1.01 (0.99-1.04) in mild cognitive impairment (interaction p = 0.013). C-index gain +0.02 in subjective cognitive decline; cutoff of -2.6 years gave negative predictive value 0.91-0.99

Brain-predicted age difference is the gap between the age an algorithm estimates from a structural MRI and the person's actual age. It has been linked to dementia progression, and the question here is whether it adds anything to what a neuroradiologist already reports from visual rating scales.

Seven hundred and ninety-nine memory clinic patients from the Amsterdam Dementia Cohort — 412 with subjective cognitive decline, 387 with mild cognitive impairment — were followed for a median of 3.2 years, during which 235 (29.4%) progressed to dementia. Higher brain-PAD predicted progression (hazard ratio 1.04 per year of gap, 95% CI 1.02-1.06) and improved model fit beyond visual ratings.

The interaction is the finding. In subjective cognitive decline the hazard ratio was 1.09 (1.03-1.15) and brain-PAD improved on visual rating scales, with a modest gain in discrimination (C-index +0.02). In mild cognitive impairment it was 1.01 (0.99-1.04) and added nothing at all. A data-driven cutoff of -2.6 years — a brain looking younger than its owner — carried a negative predictive value of 0.91 to 0.99 over two to ten years.

That pattern makes sense: by the time impairment is measurable, the structural signal is already visible to a reader, and the algorithm has nothing left to contribute. Its potential use is upstream, in the patients who complain but test normally, and specifically to reassure.

• Do not order brain-age analysis; it is not a clinically available test
• The signal is in subjective cognitive decline, not in mild cognitive impairment
• A younger-than-expected brain carried a high negative predictive value — its use is reassurance
• Continue visual rating scales; brain-PAD did not replace them
• Note that brain-PAD improved fit beyond amyloid status but not discrimination

Why it matters

It locates the only stage at which an automated brain-age measure adds to what a radiologist already sees.

Don't overread it

In mild cognitive impairment, brain-PAD added nothing (hazard ratio 1.01, interval crossing 1.0); this is not a general-purpose dementia risk tool.

The statistics, in plain English

A hazard ratio of 1.04 applies per year of brain-age gap, so a patient whose brain looks five years older carries roughly a 20% higher rate, not 4%. The C-index gain of 0.02 is very small — real, but not something that changes an individual decision. The negative predictive value of 0.91 to 0.99 is high partly because most people with subjective decline do not progress anyway; high negative predictive values in low-prevalence groups are easy to achieve and easy to over-interpret.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.