Brain-wide associations between white matter and age highlight the role of fornix microstructure in brain ageing

Abstract

Identifying white matter (WM) microstructure parameters that reflect the underlying biology of the brain will advance our understanding of ageing and brain health. In this extensive comparison of brain age predictions and age-associations of WM features from different diffusion approaches, we analysed UK Biobank diffusion Magnetic Resonance Imaging (dMRI) data across midlife and older age (N = 35,749, 44.6 to 82.8 years of age). Conventional and advanced dMRI approaches were consistent in predicting brain age; with their WM-features similarly related to and predicted by age. However, brain age was estimated best when combining approaches, showing different aspects of WM to contribute to brain age. Fornix was found as the central region for brain age predictions across diffusion approaches. We encourage the application of multiple dMRI approaches for detailed insights into WM, and the further investigation of fornix as a potential biomarker of brain age and ageing.

Bio-psycho-social factors’ associations with brain age: a large-scale UK Biobank diffusion study of 35,749 participants

“…Additionally, there are differences between models trained on voxel-level compared to region-averaged data. Deep learning models using voxel-level data reach age predictions errors as low as MAE = 2.14 years in midlife to late adulthood (Peng et al, 2021) or MAE = 3.90 years across the lifespan (Leonardsen et al, 2022) while explaining large proportions of variance in age (R 2 > 0.90), whereas models trained on regional and global average measures predict age usually with larger error, MAE > 3.6 years, and/or lower variances explained R 2 < 0.75 (Beck et al, 2021, 2022; de Lange et al, 2020a, 2020b; Korbmacher et al, 2022; Rokicki et al, 2021). However, Niu et al (2019) showed that with different shallow and deep machine learning algorithms (ridge regression, support vector regressor, Gaussian process regressor, deep neural networks) high prediction accuracies (R 2 > 0.75, MAE < 1.43) could be reached when using multimodal regional average data using a young sample with narrow age range.…”

“…Brain age was predicted using the XGBoost algorithm implemented in Python (v3.7.1), being a highly effective algorithm for tabular data (Chen & Guestrin, 2016). From the total included sample (N = 35,749), we used 10% (N = 3,575) for hyperparameter tuning on a data set containing data from all diffusion approaches (i.e., full multimodal data with 1,932 features/parameters) using 5-fold cross-validation (after estimating an optimal hyperparameter tuning set size (Korbmacher et al, 2022)). The considered hyperparameters for the randomized grid search were learning rate with a range of 0.01 to 0.3 and steps of 0.05, maximum layers/depth with a range of 3 to 6 and steps of 1, and number of trees with a range of 100 to 1000 and steps of 50.…”

“…The resulting hyperparameters (learning rate = 0.05, max layers/depth = 3, and the number of trees = 750) were then used in a 10-fold cross-validation applied to the test set (N = 32,174). Cross-validaton was used to leverage the full sample size and to calculate the uncertainty around the estimates (for such see Korbmacher et al, 2022). The cross-validation procedure was executed using each of the six diffusion approaches’ metrics, whole-brain averaged metrics for all approaches (mean multimodal model), and finally a combination of all approaches and the whole-brain average scores (full multimodal model), resulting in eight brain age models (see Supplementary Table 1 for dMRI approach-specific metrics).…”

“…In other words, different from 2.4.1, we applied one model per bio-psycho-social factor. For simplicity, this analysis step only considered the best brain age predictions from the multimodal model including the metrics of all diffusion approaches (Korbmacher et al, 2022). …”

“…Furthermore, dMRI offers both single and multi-shell approaches and various meaningful metrics describing white matter microstructure (Fieremans et al, 2011; Jensen et al, 2005; Kaden et al, 2016a, 2016b; Reisert et al, 2017), which serve as a good basis for brain age estimations (Beck et al, 2021; Korbmacher et al, 2022) exploiting biophysically meaningful parameters of brain tissue in contrast to general measures such as grey/white volume or thickness. Differences in brain age-phenotype relationships can be expected when varying dMRI approaches, as varying underlying dMRI approaches will also produce variability in brain age predictions (see Beck et al, 2021; Korbmacher et al, 2022), potentially due to measuring different bio-physical processes (Fieremans et al, 2011; Jensen et al, 2005; Kaden et al, 2016a, 2016b; Reisert et al, 2017). These potential differences become important when attempting to generalize findings on brain age across the literature and setting standards for brain age predictions.…”