Multimodal brain age prediction reveals dissociable signatures of health, cognition and disease risk in 24,648 UK Biobank participants

Deep-learning brain-age models usually rely on a single MRI contrast, although brain aging affects grey matter, white matter, iron-rich nuclei and cerebrovascular tissue. Here we trained 3D DenseNet121 models to predict chronological age from five MRI modalities - T1, T2 FLAIR, T1+T2 early fusion, diffusion MRI and susceptibility-weighted imaging - in up to 24,648 UK Biobank participants, with external validation in the Parkinson’s Progression Markers Initiative. T1+T2 fusion achieved the lowest error (mean absolute error 2.19 yr; Pearson r = 0.934), but downstream analyses showed modality-specific signals. Diffusion MRI brain age gap was uniquely associated with arterial stiffness and most strongly predicted incident type 2 diabetes; susceptibility-weighted imaging showed the largest cognitive effect sizes; T2 FLAIR best predicted all-cause dementia and cerebrovascular disease; and diffusion MRI carried the highest hazard for Alzheimer’s disease. Grad-CAM indicated that fusion shifted attribution from grey to white matter, and Mendelian randomization implicated physical activity, diet, smoking and alcohol intake as causal lifestyle determinants. These findings establish brain age as a family of modality-specific biomarkers.