AI-powered Surreal-GAN identifies diverse brain aging patterns

By Dr. Sushama R. Chaphalkar, PhD.Reviewed by Benedette Cuffari, M.Sc.Aug 20 2024

In a recent study published in Nature Medicine, researchers utilize artificial intelligence (AI)-powered ‘Surreal-GAN,’ a deep-representation learning model, to examine the heterogeneity of aging brains.

Study: Brain aging patterns in a large and diverse cohort of 49,482 individuals. Image Credit: Wirestock Creators / Shutterstock.com

The neurology of aging

The aging human brain undergoes many structural changes that vary based on genetics, an individual’s lifestyle, and the presence of coexisting diseases. This heterogeneity in brain aging is also influenced by factors that can either exacerbate or protect against age-related neuropathological processes.

Early and subtle changes in specific brain regions can emerge during the preclinical phases of neurodegenerative diseases like Alzheimer’s disease. Therefore, it is crucial to understand these neuroanatomical changes across a broad spectrum of individuals.

Traditional neuroimaging studies have provided important insights into the role of aging and diseases in altering brain structure and function, often relying on case-control comparisons. However, these methods are limited in their ability to address individual heterogeneity, as they typically focus on average patterns, rather than capturing the diverse neuroanatomical changes across individuals.

Machine learning (ML) approaches have been successfully used to identify individual-level neuroimaging biomarkers of brain aging. However, these methods often do not consider underlying heterogeneity and, as a result, only identify biomarkers that reflect a typical or average pattern.

About the study

The present study used Surreal-GAN to analyze brain aging patterns by learning one-to-many transformations from a reference (REF) population to a target (TAR) population to capture heterogeneous brain changes. These data are then distilled into representation (R) indices that quantify the severity of individualized brain changes along several dimensions.

The REF group comprised 1,150 participants between 20 and 49 years of age, whereas the TAR group included 8,992 individuals between 50 and 97 years of age, including those with mild cognitive impairment (MCI) or dementia.

To ensure robustness, the model's reproducibility was tested on an independent training set and across sexes. The researchers also associated R-indices with demographic, clinical, neurocognitive, lifestyle, and genetic measures using partial correlations, voxel-based morphometry, and genome-wide association studies (GWAS).

Chronic disease associations were explored through multiple linear regression analysis that adjusted for both age and sex. Prognostic capabilities were assessed using Cox proportional hazard models on longitudinal data.

Study findings

Five distinct R-indices of R1, R2, R3, R4, and R5 represented subcortical, medial temporal lobe, parieto-temporal, diffuse cortical, and perisylvian atrophy, respectively. These indices were significantly associated with demographic variables, such as age and sex, as well as several chronic diseases, including MCI, dementia, schizophrenia, and Parkinson’s disease.

R2, R3, and R5 were strongly correlated with Alzheimer’s disease biomarkers, particularly cerebrospinal fluid (CSF)-pTau181 and CSF-Aβ42. R5 was associated with a wide range of chronic diseases and lifestyle factors, such as alcohol intake and smoking.

Baseline R-indices were strong predictors of disease progression from cognitively normal to MCI and from MCI to dementia, as well as mortality risk.

Several genetic loci were associated with these R-indices, including some that were not previously linked to clinical traits. Thus, R-indices can both capture the heterogeneity of brain aging and serve as valuable markers for understanding the progression of neurodegenerative diseases, as well as the impact of lifestyle and genetic factors on brain health.

The current study is strengthened by the use of an enhanced Surreal-GAN methodology with broad applicability for discovering patterns of brain aging, thereby offering an expandable system applicable to diverse research questions. However, notable limitations of the current study include the underrepresentation of uncommon pathologies and certain diseases, as well as by the chosen age threshold, which may obscure some preclinical changes in the REF group.

Conclusions

The researchers developed Surreal-GAN, a five-dimensional representation system to characterize the neuroanatomical heterogeneity of brain aging. Surreal-GAN provides a novel tool for dissecting brain atrophy heterogeneity and understanding its links to demographic, pathological, lifestyle factors, and genetic variants. Furthermore, Surreal-GAN also has the potential to enhance personalized diagnostics, patient management, and improve the precision and effectiveness of clinical trials.