| dc.description.abstract | Introduction: Examining brain changes through a lifespan perspective allows for multiple ways of answering fundamental questions about the brain’s development and to identify biomarkers of ageing and disease. One opportunity is to establish a connection between the brain’s architecture and chronological age using brain magnetic resonance imaging (MRI) features. The resulting metric, often called brain age, can be used as a proxy of a person’s general health status when comparing the prediction to the chronological age, and can hence be used as a general identifier of disease or symptom load. An adjacent goal was to establish a better understanding of the brain’s development and ageing by examining MRI-features age dependencies, such as white and grey matter. Here, we provide an overview of five studies conducted during this PhD project, examining brain ageing as well as ageing and disease biomarkers, and contextualise these in the embedding of the wider research field.
Methods: We examined brain-wide and hemispheric data from the voxel-to-globallevel in large cohorts and precision imaging data of healthy controls with a focus on age-associations. We used several data sets, yet mainly cross-sectional MRI data from the UK Biobank (N ⇡ 50,000). Additionally, we used longitudinal UK Biobank data incorporating two time-points (N = 2,678), local cross-sectional data (Tematisk Område Psykoser (TOP) and Neurogenetics of cognition in aging (NCNG), Ntotal = 1,065), and repeatedly sampled data (Bergen Breakfast Scanning Club (BBSC) and Frequently Travelling Human Phantom (FTHP)) Ntotal = 4 (M = 660 scans). For one paper (Study C) we used minimally processed T1-weighted 3D volumes, which were aligned in MNI space and empty-border cropped. All other studies used processed and regionally or globally atlas-averaged white matter (WM) and grey matter (GM) metrics (John Hopkins University WM atlas, and Desikan-Killiany atlas for GM and WM thickness, area, and volume, respectively). Moreover, after estimating WM microstructure metrics based on different biophysical models, we used tract-based spatial statistics to obtain diffusion metrics of conventional and advanced diffusion approaches within the fractional anisotropy (WM) skeleton. In brief, these approaches contain radial, axial and mean diffusivity metrics, fractional anisotropy and water fractions, kurtosis metrics, and additional information for these metrics for intra- and extra-axonal space. For brain age estimations, we applied both a tree-based algorithm (extreme gradient boosting) to processed and region-averaged multimodal magnetic resonance imaging metrics (2D), and a convolutional neural network (CNN) to minimally processed structural MRI data (3D).
Statistical analyses included associating the estimated brain ages with each other and chronological age. All regional and global WM and GM metrics and their asymmetries were age-associated for both cross-sectional and longitudinal data. We also associated the resulting WM brain age with various bio-psycho-social phenotypes, and assessed cross-sectional and longitudinal measures WM measures with polygenic risk scores.
Finally, we tested an existing brain age model’s applicability in practical, clinical context (a pre-trained CNN using T1-weighted MRI data) by assessing the models prediction error in repeatedly sampled data of few individuals and in a cross-sectional validation sample, as well as associations of brain age predictions with quality control metrics and field strength.
Results: We benchmarked age-associations of various GM and WM metrics in midlife to older ages and examined brain age estimated from varying MRI data in different samples. Global brain ageing was characterised by decreases in GM volume, surface area, and thickness, and WM fractional anisotropy, the intra-axonal water fraction, and kurtosis metrics decreases, an axial, radial, and mean diffusivity metrics as well as free water and extra-axonal water fractions increases. These trends were similar for regional age-associations and were presented as age charts of normative white and grey matter development across mid- and late life.
WM and GM regions presented some variability for age estimations, with multimodal models presenting most accurate predictions. Yet, brain ages estimated on metrics from different biophysical models were similarly associated to different biopsycho-social factors. Moreover, models including diffusion MRI derived WM metrics provided slightly more accurate brain age estimates than T1-weighted MRI derived metrics.
Across studies, the fornix repeatedly appeared as one of the most age-sensitive region across WM metrics in the UK Biobank. However also forceps minor was highly age-sensitive in addition to the middle cerebral peduncle, which was strongest related to the polygenic risk of Alzheimer’s Disease. The annual rate of change in WM was a magnitude stronger associated to the polygenic risk of Alzheimer’s Disease and psychiatric disorders than cross-sectional measures. However, effect sizes were small for both global and regional brain-polygenic risk associations.
We furthermore identified a tendency of higher regional GM and WM asymmetry at higher ages. In that sense, for GM, amygdala, hippocampus, pallidum, ventricle volume, thalamus and accumbens were strongest associated with age. For WM, the cingulate tract, unicate fasciculus, superior longitudinal fasciculus, and cerebral peduncle asymmetries were strongest associated with age. New measures of hemispherespecific age-predictions were suggested and demonstrated promising results to further investigate asymmetric ageing of the brain or diseases affecting the brain asymmetrically. Predictions were highly similar across hemispheres, modalities and handednesspreference. Yet, important sex-differences applied to the sensitivity of brain age to the MRI modality and hemisphere.
When testing a pre-trained brain age model in repeatedly/densely sampled data, we found low correlations between predicted and chronological age in the repeatedly sampled data. In an attempt of explaining such variability by scan quality, we identified inconsistent associations of brain age with quality control parameters. We also found a stronger associations of brain and chronological ages at a higher field strength for one repeatedly sampled individual (where data was available at different field strengths), which was validated in cross-sectional samples.
Conclusion: Central and deep brain regions including the corpus callosum, the brain stem, the limbic system, and the ventricles were regions which were repeatedly strongly associated with age, ageing, and the polygenic risk of ageing related pathology. Among these regions, fornix stood out as most prominent region. Fornix characteristics have been identified previously as biomarkers of Alzheimer’s Disease (AD) progression. Subsequently, fornix has already been used as a target region for deep brain stimulation for AD treatment. We underlined these findings by showing the region’s strong age-association, which might be useful to identify earlier stages of cognitive decline and neurodegenerative disease. Additionally, longitudinal changes unveiled a unique pattern of WM changes not only affecting the limbic system (including fornix) but also presenting an anterior-posterior gradient of WM change of WM loss and dedifferentiation in superior frontal regions compared to differentiating and potentially plasticity in occipital, brain stem and cerebellar regions. We furthermore provided a spatially distributed pattern of associations between polygenic risk scores and WM, particularly outlining the cerebral peduncle. These genetically informed risk scores associations with brain WM were stronger for the annual change in WM than timepoint specific/cross-sectional scores, emphasising the importance of using longitudinal data. The newly tested hemispheric brain age might hold some promise for precision medicine by assessing differences between left and right brain ages. Concerning the test of the existing model, part of the prediction error seems to be driven by differences in field strength. Yet, additional confounds need to be identified and addressed to move brain age towards higher clinical utility. | en_US |
