Researchers looked for the brain patterns behind changes in the “brain age gap” — a machine‑learning estimate of how much a brain looks older than a person’s real age — in people with mild cognitive impairment (MCI). They used a model called a coVariance neural network (VNN) that processes brain scans by projecting them onto the principal patterns of anatomical covariance. In simple terms, those principal patterns (called covariance eigenvectors) are coordinated patterns of cortical thickness across brain regions. The team asked whether a longitudinal change in brain‑age gap between amyloid‑positive and amyloid‑negative people could be traced to those latent patterns instead of to single brain regions.

What the researchers did: they trained and used VNNs on cortical thickness data and measured brain‑age gap over time in 464 MCI participants from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). To connect model predictions to brain measurements they used Integrated Gradients, an attribution method that divides a prediction difference among input directions. They combined those per‑scan attributions with a fitted linear mixed model (which adjusts for participant differences and weights repeated visits) so the contribution of each latent direction or brain region summed to the reported group‑by‑time effect.

What they found: one leading covariance eigenvector — a single coordinated pattern across brain regions — accounts for 83.2% of the adjusted difference in annual brain‑age gap change between amyloid‑positive and amyloid‑negative MCI groups. By contrast, no single brain region explained more than 7.7% of that effect. When they artificially kept or removed that dominant direction from the VNN inputs, the model’s longitudinal effect changed in the way predicted by the attribution. The result was robust to resampling participants, different training starting points, and changes in preprocessing.