New AI Models Predict Biological Age with Brain Scans and Metabolomic

Recent research has leveraged federated learning to enhance the prediction of biological age using brain MRI images and metabolomic biomarkers. By integrating data from multiple cohorts, the study demonstrated how biological age scores such as BrainAge and MetaboAge provide insights into different aspects of aging, with potential applications for mortality prediction and health assessments.

What’s New:

The study introduces a federated deep learning model to estimate BrainAge, a biological age score based on brain MRI scans. By training this model across three large cohorts, the researchers achieved greater accuracy in age prediction than conventional, locally trained models. Additionally, the study examined MetaboAge, a metabolomic-based biological age score, comparing it with BrainAge to determine their predictive value for aging-related outcomes. The combined use of both scores showed a higher predictive value for mortality than either score alone, indicating they capture different dimensions of the aging process.

How It Works:

Biological age scores are regression models that estimate a person’s age based on physiological biomarkers rather than just chronological age. These scores are valuable for identifying health risks associated with aging, as biological and chronological aging rates often differ significantly between individuals. BrainAge, for example, is derived from MRI data that reflects structural changes in the brain associated with aging, while MetaboAge is based on metabolomic biomarkers from blood samples that indicate metabolic health.

The research utilized federated learning, a machine learning approach that allows different institutions to collaborate on training a model without sharing sensitive data. Instead, the model is trained locally at each site, and only aggregated results are shared. This method not only enhances privacy but also improves model generalizability across diverse populations. The federated BrainAge model was able to predict age more accurately across multiple cohorts, reducing error rates compared to locally trained models.

Behind the News:

The study builds on previous work that has used brain imaging and metabolomics to estimate biological age. Earlier approaches to calculating BrainAge or MetaboAge were limited by smaller datasets and locally trained models, which often resulted in overfitting and poor generalization to new populations. Federated learning addresses these issues by allowing large-scale collaboration without the need for data sharing, leading to more robust and accurate models.

The researchers trained the BrainAge model using three population-based cohorts: the Rotterdam Study (RS), the Maastricht Study (TMS), and the Leiden Longevity Study (LLS). They found that the federated model outperformed local models in predicting chronological age with mean absolute errors (MAE) ranging from 4.36 to 5.59 years. Adjusting for bias in different age groups further improved the model's accuracy, particularly in older cohorts.

Why It Matters:

The combination of BrainAge and MetaboAge offers a more comprehensive view of aging, as each score reflects different biological processes. The study found a small but significant association between the two scores, likely due to their shared correlation with chronological age. However, when used together, they provided a more powerful prediction for mortality, suggesting they capture complementary aspects of aging.

Malfunctioning brain structure and metabolic processes are key indicators of various age-related diseases, including dementia, cardiovascular issues, and general mortality risks. By incorporating both brain imaging and metabolomic data, healthcare providers could gain a deeper understanding of a patient’s health status beyond chronological age, improving the ability to detect early signs of disease.

We’re Thinking:

The use of federated learning in aging research marks a significant advancement, particularly for improving model accuracy while ensuring data privacy. The dual approach of combining BrainAge and MetaboAge may also lead to better health monitoring systems that can identify high-risk individuals more effectively, paving the way for personalized treatment plans and earlier interventions.

Challenges and Future Directions:

While the study shows promising results, there are challenges in generalizing the findings to other populations. The cohorts used were predominantly of Western European descent, and the models might need further validation in more diverse populations. Additionally, the federated learning process, while improving privacy, requires careful coordination and harmonization across institutions, which can be resource-intensive.

The next step for this research could involve expanding the use of combined biological age measures in clinical settings to refine predictions for aging-related diseases. Further exploration of these models in diverse demographic groups could also enhance their applicability and effectiveness in global health.

The study has been published on Arxiv.