Multi-Cohort Federated Learning Shows Synergy in Mortality Prediction for MRI-Based and Metabolomics-Based Age Scores

Area of Science:

• Biomedical data science
• Aging research
• Neuroimaging and metabolomics

Background:

• Biological age scores estimate physiological aging but their interactions are poorly understood.
• Large-scale multi-modal data are essential for studying these interactions, yet data sharing is restricted.
• Federated learning offers a solution for analyzing distributed, sensitive health data.

Purpose of the Study:

• To investigate the relationship between BrainAge (from brain MRI) and MetaboAge (from metabolites) using federated learning.
• To assess the predictive performance of a federated deep learning model for BrainAge.
• To evaluate the complementary roles of BrainAge and MetaboAge in predicting dementia and mortality.

Main Methods:

• Employed federated learning across three large population-based cohorts to train a deep learning model for BrainAge estimation.
• Compared the federated model's performance against models trained within single cohorts.
• Conducted association and survival analyses to compare BrainAge and MetaboAge for dementia and mortality prediction.

Main Results:

• The federated BrainAge model significantly reduced age prediction error compared to local models.
• Harmonizing age intervals further enhanced federated BrainAge accuracy.
• BrainAge and MetaboAge exhibited a weak association but complementary predictive values for mortality risk.

Conclusions:

• Federated learning is effective for analyzing restricted research cohort data.
• BrainAge and MetaboAge synergistically predict all-cause mortality risk.
• These distinct biological age scores capture different facets of the aging process.