Introduction: Brain-age difference has been suggested as a biomarker of healthy aging and a method of capturing brain state (see also, relative brain age (RBA), brain-predicted age difference (brainPAD), brain age gap estimate (brainAGE), etc.). Associations have been found between brain-age differences and lifestyle variables (e.g., body mass index (BMI) and smoking status), and between measures of brain-age difference and cognition in cognitively impaired populations. However, few studies have identified relationships between brain-age differences and age-related outcomes, such as cognition, in healthy adults. Methods used to calculate brain-age differences have proliferated; many studies suggest brain-age difference methods that vary in prediction model, input features, and training population. It therefore remains unclear whether brain-age differences have utility in healthy adults and whether this utility varies with the method of brain-age difference calculation.
Methods: In the current study, three brain-age difference methods (Elastic Net regression, 3D Convolutional Neural Network, and Random Forest regression), each using measures of structural brain state based on T1-MPRAGE images, were compared. The Elastic Net regression and 3D Convolutional Neural Network models have been trained in independent studies on separate datasets using gray matter density (Elastic Net; n = 1,359, mean age (SD) = 40.04 (17.78) years) and voxel-wise and surface-based volume (3D Convolutional Neural Network; n = 5,851, mean age (SD) = 56.74 (16.88) years) and used to assess cognition. Random Forest regression, on the other hand, was trained for the current study on structural morphometry, including volume, thickness, and surface area, from the Human Connectome Project-Aging sample (n = 660, mean age (SD) = 59.04 (14.63) years). Each of these trained models were used to compute brain-age differences from a single, independent sample of healthy adults aged 36-87 years old (n = 99, mean age (SD) = 61.03 (10.79) years). Models were compared by assessing model fit, brain-age differences, and feature contribution. Use of brain-age differences as a biomarker of healthy aging was assessed as the relationship between brain-age differences and five cognitive domains (episodic memory, executive function, processing speed, language, and mental status). Further, the moderating role of contextual features relevant to brain health—chronological age and cardiorespiratory fitness—on the relationship between brain-age differences and cognition was assessed.
Results: Each brain-age difference model had different levels of age-prediction accuracy and had distinct patterns of input features that contributed to predictions, despite sharing some important input features (inferior parietal, temporal, and medial orbitofrontal cortices. Brain-age differences from Elastic Net regression were associated with cognition including processing speed and mental status. Cardiorespiratory fitness also moderated the relationship between Elastic Net brain-age differences and processing speed, such that greater cardiorespiratory fitness was associated with a weaker relationship between brain-age difference and processing speed. However, no relationships were found between any cognitive domains and brain-age differences using 3D Convolutional Neural Network and Random Forest models.
Discussion: These findings suggest that brain-age differences vary with methodology, that these variations in brain-age differences can influence results, and that these differences may influence the utility of brain-age difference as a biomarker of the structural brain and cognitive health in healthy aging.