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

Alzheimer’s Disease (AD) remains a critical public health challenge due to its late clinical onset and lack of effective early interventions. Brain age prediction provides a powerful framework to detect deviations from normative aging and identify individuals at elevated risk. We present a multimodal, interpretable deep learning approach using graph neural networks (GNNs) to estimate the brain age gap (BAG)—the difference between predicted and chronological age. Structural and functional brain connectivity were modeled as graphs, with nodes encoding regional microstructure and volume, and edges derived from diffusion and/or resting-state functional MRI. Additional features included APOE genotype, transcriptomic components, demographic factors, and fluid biomarkers. SHAP (SHapley Additive exPlanations) values enabled regional attribution and biologically informed contrastive learning. Across three datasets—ADNI, AD-DECODE, and Duke-UNC ADRC—our models achieved strong predictive accuracy: ADNI (MAE: 3.78 ± 0.68 years; R^2: 0.25 ± 0.24), AD-DECODE (MAE: 6.34 ± 1.62; R^2: 0.71 ± 0.14), and ADRC (MAE: 7.15 ± 2.5; R^2: 0.36 ± 0.26). Higher corrected BAG values were associated with atrophy in hippocampal, amygdalar, sensorimotor, cerebellar, and basal ganglia regions, with hippocampal effects especially evident longitudinally. SHAP-based clustering revealed biologically meaningful subgroups linked to molecular profiles, and BAG correlated with CSF pTau, Aβ42/40, and transcriptomic signals (e.g., PSEN1, CCL3). This interpretable framework enables biologically grounded predictions and subgroup discovery, enhancing the clinical translation of AI models in AD research.

Duke Scholars

Author Alexandra Badea Radiology