Novel Framework for Brain Age Prediction using Graph Neural Networks

Abstract

AbstractBackgroundAging is among the largest risk factors for neurodegeneration which can be accelerated by age‐related pathologies. Hence, individuals may age at different rates reflected by a gap in biological estimates of age and actual chronological age. Graph neural networks (GNN) are deep learning models that leverage spatial information inherent in the dataset for inference. Here, we develop an MRI cortical thickness (CT) based brain age prediction framework using GNNs and illustrate that it provides a feasible mechanism to identify contributing regions leading to elevated brain age gap in Alzheimer’s disease (AD).MethodWe used ANTs to calculate CT in 100 parcels from 3T MRI of 170 healthy controls (HC; age = 64.26±8.27y, 101 females), 53 mild cognitive impairment (MCI; age = 68.56±8.58y, 22 females), and 62 AD (age = 67.25±8.83y, 30 females) individuals. A GNN model based on the anatomical covariance matrix from HC cohort was trained to predict chronological age from associated CT for HC cohort. GNN output was evaluated as an unweighted mean of entities associated with the brain regions in the final layer. Hence, individual contributions to the GNN output by each brain region could be quantified. Final brain age prediction was obtained after removing age‐bias from GNN outputs using linear regression model. For all subjects, we evaluated residuals at regional level (regional age‐gap) by comparing the contribution of each parcel with corresponding GNN output. The residual between predicted brain age and chronological age (age‐gap) was also evaluated. See Figure 1 for an overview.ResultANCOVA (with correction for age and gender) with post‐hoc analyses revealed that regional age‐gap was greater for AD relative to HC in inferior parietal, temporal pole, inferior temporal, medial temporal, and superior frontal regions (Bonferroni corrected p‐value<0.05). Also, regional age‐gap was greater for MCI relative to HC for inferior parietal and temporal pole regions (Figure 2). Subsequently, age‐gap was elevated in AD (mean = 5.64y) and MCI (mean = 2.66y) relative to HC (mean = 0y) reflecting a linear increase in age‐gap with disease stage (Figure 3).ConclusionOur novel framework identified regional CT associations with age‐gap in AD and MCI, suggesting that this GNN approach may be used to further our understanding between aging and neurodegeneration.