Multimodal graph attention networks for predicting brain age and stratifying Alzheimer’s disease risk in clinical populations

Alzheimer’s Disease (AD) remains a major public health challenge due to its late clinical onset and limited options for early intervention. Brain age prediction quantifies deviations from normative aging to identify individuals at risk. We present a scalable, interpretable graph-based deep learning framework that integrates multimodal neuroimaging and biological data to estimate the brain age gap (BAG). Using Graph Neural Networks (GNNs), we modeled the brain as a network of regions with edges derived from diffusion and/or resting-state functional MRI connectomes. Node features captured regional microstructure and volume, while auxiliary inputs included clinical and demographic variables, APOE genotype, transcriptomic components, and fluid biomarkers. The architecture—based on Graph Attention Networks (GATs)—was pretrained on healthy individuals and fine-tuned via transfer learning. Interpretability was achieved using SHAP (SHapley Additive exPlanations) values, which also supported biologically informed contrastive learning. Across three independent cohorts (ADNI, AD-DECODE, and Duke-UNC ADRC), our models achieved strong predictive performance (MAE ~5.7–5.9 years; R² ~0.67–0.69). Higher cBAG values were significantly associated with regional atrophy in hippocampal, amygdalar, sensorimotor, cerebellar, and basal ganglia structures, with hippocampal effects specific to longitudinal trajectories. SHAP-based clustering identified biologically meaningful subgroups based on regional contributions and molecular profiles. Notably, cBAG correlated with fluid biomarkers (CSF pTau, Aβ42/40) and transcriptomic signals (e.g., PSEN1, CCL3), linking predicted brain aging to established AD pathology. This interpretable multimodal framework advances precision neuroimaging by supporting subject stratification and mechanistic insight— key components for clinical translation of AI in Alzheimer’s research.