Study: Brain clock captures variation and disparities in aging and dementia in geographically diverse populations. Image credit: Lightspring / Shutterstock
A recent study published in the journal Nature Medicine, Researchers used deep learning to analyze the impact of geographic, socio-demographic, socio-economic, neurodegeneration-related, and gender diversity on brain age disparities in 15 countries. They found that structural socio-economic inequalities, pollution, and health disparities were significant predictors of widening brain age disparities, particularly in the Latin America and Caribbean (LAC) region, with larger disparities observed in women and individuals with cognitive impairments, such as Alzheimer's disease (AD).
background
The brain changes dynamically with age, which is important to understand, especially in the context of disparities and brain disorders such as Alzheimer's disease. Brain age models that measure brain health across multiple factors have the potential to capture diversity in aging, but are understudied in underrepresented populations such as LAC. These populations face significant socioeconomic and health disparities that may impact brain aging. Research on brain aging has focused primarily on northern hemisphere populations and often uses structural magnetic resonance imaging (MRI), ignoring brain network dynamics captured with functional MRI (fMRI) or electroencephalography (EEG). While EEG is a more accessible tool in resource-limited settings, its use in large-scale studies is limited by challenges in standardization and integration with fMRI. There is a need to develop scalable brain age markers using deep learning that incorporate these techniques and account for demographic diversity, especially in underrepresented populations. Therefore, the researchers in this study used graph convolutional networks to predict disparities in brain age and explore the impact of geographic, sociodemographic, and health-related variability on brain aging.
About the Research
The study analyzed resting-state fMRI and EEG datasets from 5,306 participants across 15 countries in LAC and non-LAC regions. fMRI data were collected from 2,953 participants in Argentina, Chile, Colombia, Mexico, Peru, the United States, China, and Japan, while EEG data were collected from 2,353 participants in Argentina, Greece, Brazil, Chile, Colombia, Cuba, Ireland, Italy, Turkey, and the United Kingdom. Participants included 3,509 healthy controls and 1,808 individuals with neurocognitive disorders such as mild cognitive impairment (MCI), AD, or behavioral variant frontotemporal dementia (bvFTD). Data underwent rigorous preprocessing, including normalization, noise correction, and source space estimation. Higher-order interactions between brain regions were assessed, and data were graph-transformed for analysis via graph convolutional networks (GCN). An approach involving 80% cross-validation and 20% holdout testing was used. Data augmentation techniques were employed and the predictive performance of the model was evaluated using goodness of fit (R²) and root mean square error (rmse). Gradient boosting models were used to explore the influence of exposome factors on the brain age gap. Extensive statistical analyses, including permutation tests and bootstrapping, were conducted to validate the results. The quality of the data was carefully assessed and the study adhered to strict ethical guidelines.
The datasets included LAC and non-LAC healthy controls (HC, total n = 3,509) and participants with Alzheimer's disease (AD, total n = 828), bvFTD (total n = 463), and MCI (total n = 517). The fMRI dataset included 2,953 participants from LAC (Argentina, Chile, Colombia, Mexico, and Peru) and non-LAC (US, China, and Japan). The EEG dataset included 2,353 participants from Argentina, Brazil, Chile, Colombia, and Cuba (LAC) and Greece, Ireland, Italy, Turkey, and the UK (non-LAC). Raw fMRI and EEG signals were preprocessed by filtering and artifact removal, and EEG signals were normalized to project into source space. Segmentation using the Automatic Anatomical Labeling (AAL) atlas was performed for both fMRI and EEG signals and nodes were constructed to compute higher order interactions using the Ω information metric. Connectivity matrices for both modalities were obtained and later represented in a graph. Data augmentation was performed only on the test dataset. The graph was used as input for a graph convolutional deep learning network (architecture shown in the last row) with separate models for EEG and fMRI. Finally, age predictions were obtained and performance was measured by comparing predicted age with real age. This figure was partially created using BioRender.com (fMRI and EEG devices).
Results and Discussion
The brain aging model showed adequate predictive performance. Key predictive brain region features were centered in a fronto-posterior network, including nodes in the precentral gyrus, middle occipital gyrus, superior frontal gyrus and middle frontal gyrus. Other key nodes in the fMRI model were the inferior frontal gyrus, anterior cingulate gyrus and middle cingulate gyrus, and paracingulate gyrus. In the EEG model, the inferior occipital gyrus and superior and inferior parietal gyrus were also important.
Notably, when analyzing the non-LAC dataset, the model showed a similar pattern in predictive capabilities, but with a slight decrease in fit. In contrast, models trained on the LAC dataset showed moderate fit and increased rmse values, highlighting a bias towards predicting older brain ages, especially in female participants. Furthermore, an examination of the between-region effect revealed that training on non-LAC data and testing on LAC resulted in a positive mean directional error (MDE), indicating a bias towards older brain ages. Furthermore, a widening gap in brain ages was observed in clinical populations, suggesting accelerated aging in conditions such as MCI and AD compared to healthy controls.
These findings highlight the complexity of brain aging in different populations and the importance of considering diversity factors in neurocognitive assessments. This study is strengthened by using a diverse dataset across multiple countries, integrating fMRI and EEG data, and developing a scalable, individualized brain health index applicable to diverse and underrepresented populations. However, this study has limitations including a lack of clinical EEG data from regions outside of LAC, reliance on a unimodal brain age gap measure, limited regional data, and a lack of individual-level demographic factors such as gender identity, socioeconomic status, and ethnicity.
Conclusion
In conclusion, this study shows that despite data variability, the brain clock model is sensitive to various factors including geography, gender, macrosocial influences, and disease. By leveraging deep learning on higher-order brain interactions across fMRI and EEG, this study paves the way for comprehensive and accessible tools to assess differential brain aging. This could aid in identifying and staging neurocognitive disorders such as MCI, AD, and bvFTD, and support personalized medicine approaches worldwide.