Accelerated MRI-predicted brain ageing and its associations with cardiometabolic and brain disorders

Arinbjörn Kolbeinsson, Sarah Filippi, Yannis Panagakis, Paul M. Matthews, Paul Elliott, Abbas Dehghan, Ioanna TzoulakiScientific 2020 · DOI

Abstract

Brain structure in later life reflects both influences of intrinsic aging and those of lifestyle, environment and disease. We developed a deep neural network model trained on brain MRI scans of healthy people to predict "healthy" brain age. Brain regions most informative for the prediction included the cerebellum, hippocampus, amygdala and insular cortex. We then applied this model to data from an independent group of people not stratified for health. A phenome-wide association analysis of over 1,410 traits in the UK Biobank with differences between the predicted and chronological ages for the second group identified significant associations with over 40 traits including diseases (e.g., type I and type II diabetes), disease risk factors (e.g., increased diastolic blood pressure and body mass index), and poorer cognitive function. These observations highlight relationships between brain and systemic health and have implications for understanding contributions of the latter to late life dementia risk.

Introduction

While chronological age is a constant, biological aging varies significantly across individuals. The brain, in particular, serves as a critical indicator of overall health and neurodegenerative risk. "Brain age"—the estimated biological state of the brain compared to an individual's actual age—has emerged as a powerful biomarker. However, previous modeling attempts have often been limited by small datasets, linear methodologies that fail to capture complex structural nuances, or training on "unselected" populations that include both healthy and diseased individuals, thereby blurring the definition of a "normal" aging trajectory.

This study addresses these gaps by developing a robust, non-linear normative model of healthy brain aging. By establishing a clear baseline for healthy individuals, researchers can identify those whose brains appear "older" than expected and investigate the systemic health factors, such as cardiovascular and metabolic conditions, that drive this accelerated aging.

Methods

Model Architecture and Training

The researchers utilized a 3D Convolutional Neural Network (CNN) based on the ResNet architecture. Unlike traditional 2D models that analyze individual brain slices, this 3D approach captures the full spatial complexity of the brain's anatomy from T1-weighted MRI scans. To ensure the model learned a pure signal of "normal" aging, it was trained exclusively on a "healthy" subset of the population (3,067 individuals) with no history of chronic disease or relevant clinical diagnoses.

Analytical Framework

To interpret the model and its clinical relevance, the team employed several advanced techniques:

  • Permutation Importance: The researchers systematically removed information from 139 different brain regions to identify which areas were most critical for accurate age prediction.
  • Phenome-wide Association Study (PheWAS): The "brain age delta" (the difference between predicted and chronological age) was tested against over 1,410 clinical and lifestyle traits in the UK Biobank.
  • Mendelian Randomization (MR): Genetic data were used to investigate whether observed correlations, such as those between blood pressure and brain age, represented a causal relationship.

Results

Model Performance

The 3D CNN demonstrated high accuracy in predicting brain age. In the healthy control group, the model achieved a Mean Absolute Error (MAE) of 2.87 years. When applied to the broader, unselected general population of 12,296 individuals, the model maintained a strong Pearson correlation (r = 0.82) between predicted and calendar age.

Correlation between predicted brain age and chronological calendar age in the general population.

Anatomical Drivers of Prediction

The permutation importance analysis revealed that while the hippocampus and amygdala were significant, the cerebellum was one of the most informative regions for the model's predictions. This finding highlights the cerebellum's underappreciated role in the aging process and its sensitivity to biological decline.

Permutation importance map highlighting the cerebellum (yellow) as a primary region for age prediction.

Clinical and Causal Associations

The PheWAS identified significant associations between accelerated brain aging and over 40 clinical traits. Notable findings include:

  • Neurological and Metabolic Disease: Multiple Sclerosis (MS) was associated with a four-fold higher risk per standard deviation of brain age delta. Both Type I and Type II diabetes were strongly linked to "older" looking brains.
  • Cognitive Function: Higher brain age deltas correlated with poorer performance on cognitive tasks, specifically symbol-digit matching.
  • Cardiovascular Health: Most significantly, the Mendelian Randomization analysis provided evidence for a causal link between diastolic blood pressure and accelerated brain aging (beta 0.06, p=0.01). This suggests that managing blood pressure could be a direct intervention to slow biological brain aging.

Discussion

This research demonstrates that deep learning models trained on healthy populations can effectively identify individuals at risk for accelerated brain aging. The causal link between diastolic blood pressure and brain atrophy is particularly salient, providing a clear target for early intervention in dementia prevention.

However, the study is not without limitations. The UK Biobank population is subject to "healthy volunteer" bias, and the cross-sectional nature of the current data limits the ability to track the rate of aging over time. Future research utilizing longitudinal scans and more diverse global cohorts will be essential to validating these findings and transitioning brain age models into routine clinical practice.

Summary updated

Abstract

Brain structure in later life reflects both influences of intrinsic aging and those of lifestyle, environment and disease. We developed a deep neural network model trained on brain MRI scans of healthy people to predict "healthy" brain age. Brain regions most informative for the prediction included the cerebellum, hippocampus, amygdala and insular cortex. We then applied this model to data from an independent group of people not stratified for health. A phenome-wide association analysis of over 1,410 traits in the UK Biobank with differences between the predicted and chronological ages for the second group identified significant associations with over 40 traits including diseases (e.g., type I and type II diabetes), disease risk factors (e.g., increased diastolic blood pressure and body mass index), and poorer cognitive function. These observations highlight relationships between brain and systemic health and have implications for understanding contributions of the latter to late life dementia risk.
BibTeX
@inproceedings{kolbeinsson2020accelerated,
  author = {Arinbjörn Kolbeinsson and Sarah Filippi and Yannis Panagakis and Paul M. Matthews and Paul Elliott and Abbas Dehghan and Ioanna Tzoulaki},
  title = {{Accelerated MRI-predicted brain ageing and its associations with cardiometabolic and brain disorders}},
  booktitle = {Scientific},
  year = {2020},
  doi = {10.1038/s41598-020-76518-z},
}
Figure