Deep Learning Model Separates Amyloid-Specific Brain Signals From Normal Aging

A contrastive deep learning analysis of brain scans from 289 older adults isolated a functional signal tied specifically to amyloid pathology, distinct from patterns shared with amyloid-negative aging brains, and linked it to attention and memory performance.

A deep learning model trained on resting state brain scans from 289 older adults has separated functional brain signals specific to amyloid pathology from those shared with ordinary aging, according to a study published August 26, 2026 in Alzheimer's and Dementia. Researchers led by Kanhao Zhao, working with the Alzheimer's Disease Neuroimaging Initiative (ADNI), classified participants as amyloid positive (129 people) or amyloid negative (160 people) using cerebrospinal fluid testing, then applied a contrastive deep learning framework to resting state fMRI connectivity data to pull apart the two kinds of signal.

The amyloid specific signal was traced mainly to the right superior temporal cortex and the anterior cingulate cortex. Using only that signal, the model predicted each person's individual cerebrospinal fluid amyloid-beta level (R-squared of 0.19, correlation of 0.46) and phosphorylated tau level (R-squared of 0.26, correlation of 0.51), tested across ten rounds of five fold cross validation within the dataset. Accuracy fell to an R-squared of 0.16 for amyloid when the same model was applied to an independent cohort, PREVENT-AD, a drop the authors report as evidence the signal generalizes but only moderately.

The regions driving these predictions were statistically associated with attention and memory test performance, the researchers reported. By contrast, the brain signal shared between amyloid positive and amyloid negative participants mapped onto language related regions and, when matched against a public gene expression atlas, aligned with molecular pathways tied to aging and vascular biology rather than Alzheimer's pathology. The amyloid specific regions, in that same gene expression analysis, aligned with pathways involving synaptic signaling and the activity of astrocytes and microglia, cell types increasingly studied as treatment targets.

The analysis is cross-sectional and confined to a single research cohort, with modest effect sizes and only partial replication in a second dataset. The authors describe their framework as a way to help interpret existing brain imaging data, not as a diagnostic tool, and the model's accuracy at predicting amyloid and tau levels in any one person remains far short of what cerebrospinal fluid testing or PET scanning already provide.

The authors say the next step is testing whether the amyloid-specific signal they isolated holds up in larger, more diverse, and longitudinally followed cohorts before any clinical application could be considered.