Brain scan models show no consistent edge in aging research
An analysis of 3,103 scans from 1,653 participants found no consistent advantage over a conventional brain activity measure for assessing current cognition.
Our read
AI & data- Studied:
- Preprint: computational study of 1,653 participants from aging cohorts
- What changed:
- Brain scan model features showed no consistent advantage over conventional brain activity measures for assessing cognition or classifying diagnoses
- The catch:
- Performance depended partly on feature extraction; different readouts improved selected cognitive associations
Editor of record: Boris Djordjevic · AgingResearch.News
Across seven aging cohorts, 1,653 participants contributed 3,103 brain scans taken at rest for an observational comparison of computer models. Two cohorts were observational; five involved interventions. The preprint tested nine configurations from five families of foundation models, systems trained on large datasets.
Features extracted from the models without further training showed no consistent advantage over conventional measures of coordinated activity between brain regions. This applied to associations with current cognitive performance and to sorting participants into diagnostic groups. Changing how researchers extracted the features improved some cognitive associations without retraining the models. Additional training and adjustments to align model outputs helped in selected settings, but not consistently across outcomes. In the main analyses of cognition over time, imaging features did not establish added value beyond initial clinical measures.
Why it matters
For smaller studies of cognitive aging, a key question is whether models trained on large brain-imaging datasets can provide useful information about cognition and its change over time.
Caveats
Although five cohorts involved interventions, the model comparison was observational, not a trial of the models’ clinical use. The preprint has not been peer-reviewed.
The paper
Stanford University
bioRxiv, Nov 9, 2026, Open access, CC BY-NC-ND, Preprint, not yet peer-reviewed
Comments
Be the first to comment.