Brain scans improved predictions of cognitive decline
Models scored 0.41 and 0.33 in one cohort and 0.42 and 0.33 in another; scores fell to 0.18–0.35 when tested across cohorts.
Our read
Biomarkers- Studied:
- Cohort study of 1,899 participants with clinical and brain imaging data (Independent replication)
- What changed:
- Adding brain scans improved machine learning predictions of cognitive decline within each research cohort
- The catch:
- The study predicted cognitive test score changes without testing if predictions improved clinical care
- In the record:
- The 2nd paper on Structural MRI with results in people; 37 on record in all.
Editor of record: Boris Djordjevic · AgingResearch.News
In 1,899 participants from two research cohorts, researchers used computational models to predict annual changes in dementia-related impairment and cognition. The cohorts included 1,237 people from the Alzheimer's Disease Neuroimaging Initiative and 662 from the Open Access Series of Imaging Studies. Participants had baseline clinical and brain-imaging data and at least two follow-up clinical assessments.
The models used demographics, health and genetic information, clinical test scores, brain-structure scans, or both. Adding scans improved predictions within each cohort. The performance scores reported above measured how much variation in impairment and cognitive change the models accounted for, respectively—not the proportion of people predicted correctly. Performance fell when researchers trained models in one cohort and tested them in the other. Models using only the 15 most predictive inputs performed similarly in external tests, suggesting that some inputs provided overlapping information.
Why it matters
Cognitive decline spans healthy age-related changes through dementia. The authors framed reliable individual predictions as a step toward more targeted early intervention and prevention.
Caveats
The models predicted changes in clinical assessment scores; the study did not test whether using these predictions improved care. External validation involved only two research cohorts.
The paper
Quantifying generalization error in machine learning prediction of cognitive decline
University of Zurich
The Journal of Prevention of Alzheimer's Disease, Aug 8, 2026, Open access

Comments
Be the first to comment.