Graph neural networks and cortical column modeling for AI-based brain age prediction in Alzheimer’s disease risk

Alzheimer’s disease (AD) affects over 10% of people above age 65. Current treatments remain largely ineffective, thus early biomarkers are essential for devising preventive interventions, and personalizing these based on risk profiles. Brain age gap (BAG)—the difference between predicted and chronological age—has emerged as a promising marker of accelerated brain aging. In this study, we estimated BAG using graph neural networks (GNNs) informed by cortical depth dependent local microstructural features derived from diffusion MRI. Brain graphs were constructed using 68 cortical regions as nodes, with edges defined by the similarity of cortical column mean diffusivity (MD) microstructural features. Cortical thickness provided nodes features. GNNs trained on MD alone achieved a mean absolute error (MAE) of 5.96 ± 1.74 years and RMSE of 7.91 ± 2.29 years (R² = 0.69 ± 0.14). Adding cortical thickness features improved performance (MAE = 5.74 ± 1.26, RMSE = 7.45 ± 1.59, R² = 0.73 ± 0.09). When applied to an APOE4-enriched cohort, the combined model achieved MAE = 4.9 and RMSE = 6.41 for brain age, and MAE = 4.41 and RMSE = 5.97 for corrected BAG (cBAG). Linear models linked cBAG to the right hippocampal volume (R² = 0.17, FDR p = 0.04). The model classified cognitive impairment with AUC = 0.80 (95% CI: 0.48–0.99). To enhance interpretability, we derived SHAP and saliency maps, identifying shared and distinct cortical contributors. Seven of the top ten regions overlapped, including the transverse temporal and entorhinal cortex. These results support cBAG as a biologically informed, personalized biomarker for early AD risk detection.