The relationship between brain‐age association and prediction: The impact of parameter selection

Abstract

AbstractBackgroundAssociation and prediction studies of brain age target the relationships between neuroanatomical changes across lifespan and behavioural or pathological phenotypes. However, the relationship between the two analyses has not been directly examined. Here we use brain resels (i.e. resolution elements) (Worsley et al. 1996) to directly compare age association and brain‐age prediction as a function of cortical thickness smoothing and parcellation parameters.MethodData included 608 healthy subjects (Age = 53.5218.07; Age range = 18‐88, 308 Female subjects) subjects with T1‐weighted MRI from the Cambridge Centre for Ageing and Neuroscience (https://www.cam‐can.org/index.php?content=dataset). CIVET pipeline (http://www.bic.mni.mcgill.ca/ServicesSoftware/CIVET) was used to extract cortical surfaces and calculate cortical thickness across the brain. We used 6 different diffusion based smoothing kernels (0, 5, 10, 20, 30, and 40 mm) and 5 different parcellation levels: 100, 200, 400, and 1000 parcels from Schaefer et al., 2018 multi‐level functional parcellation atlas as well as the entire cortex. For each level of smoothing and parcellation, age‐related correlations were calculated. Brain age was predicted using a linear regression model with 10‐fold cross validation, with principal components of cortical thickness data as predictors. To directly compare the correlation and prediction results, we calculated the brain resels for each smoothing parcellation parameter.ResultFigure 1.A shows the correlation patterns across the brain as a function of smoothing and parcellation. Figure 2 shows age prediction error as a function of smoothing kernel, parcellation level, and number of principal components included. Larger smoothing kernels and brain parcels result in higher correlation values (Figure 1.B), but lower prediction accuracy (Figure 2). Figures 3.A and 3.B show mean age‐correlation value and age‐prediction error as a function of number of resels.ConclusionOur results demonstrate an opposite relationship between brain‐age association values and brain‐age prediction accuracy, and quantify how smoothing and parcellation parameters affect each of these analyses. These results highlight the importance of parameter selection for each analysis type and how they might affect the final results, and have significant implications for brain aging studies.