GCENAN: Dynamic Edge-Conditioned Graph Attention Networks for Brain Age Estimation Leveraging Clinical Semantics

Abstract

Brain age estimation (BAE) derived from structural T1-weighted MRI has become a widely used biomarker for tracking both healthy aging and the early signs of neurodegenerative diseases, including mild cognitive impairment (MCI) and Alzheimer’s disease. Most existing deep learning models, whether based on convolutional neural networks or conventional graph neural networks, tend to handle clinico-demographic information—such as sex, socioeconomic status, MMSE, and CDR scores—as static auxiliary inputs or through simple late-fusion techniques. This common practice often overlooks the complex, subject-specific interactions between regional brain changes and an individual’s clinical profile, which in turn limits the model’s capacity to capture personalized aging trajectories and the phenomenon of regional heterochronicity. In this work, we introduce the Graph Cross-Edge-Node-Attention Network (GCENAN), a hierarchical graph neural network that builds an individualized heterogeneous graph for each subject. Within this framework, the 68 cortical regions of interest (ROIs) parcellated by the Desikan-Killiany atlas act as nodes, with their morphometric features (cortical thickness, surface area, gray matter volume, mean curvature, and Gaussian curvature) serving as node attributes. Importantly, clinico-demographic variables are incorporated as dynamic edge attributes that actively modulate the interactions between ROIs. The central innovation of our approach is the Sequential Cross-Edge-Node-Attention (GCENA) layer. This layer follows a directional two-stage updating process in which edge embeddings are first refined using multi-head cross-attention that considers both node and edge states. The updated edge representations then serve as dynamic gates for node feature aggregation. By placing clinical context at the forefront of the message-passing process, this design enables genuinely subject-specific modeling of brain aging. We evaluated GCENAN on a merged cohort of 530 subjects from the OASIS and ADNI datasets (age range 18–96 years), employing strict subject-level partitioning, 5-fold cross-validation, and a fully independent test set. The model achieved competitive performance, with a mean absolute error (MAE) of $3.12 \pm 0.45$ years on OASIS and $2.95 \pm 0.52$ years on ADNI, along with strong Pearson correlations of 0.96 and 0.95, respectively. Interpretability analyses using attention weights uncovered biologically meaningful shifts in regional importance across the lifespan, moving from motor and frontal regions in younger adults toward temporal and cingulate areas in older adults. Ablation experiments further confirmed the critical role of dynamic edge embeddings, especially cognitive markers such as MMSE and CDR, in modeling late-life brain aging. Overall, GCENAN represents a scalable and interpretable advance that goes beyond conventional static fusion methods, offering efficient inference on new subjects and richer insights into the heterogeneous nature of brain aging.