Metabolomic and lifestyle profiles refine BMI-metabolic phenotypes in older adults.

Conventional BMI-based classifications, even when combined with traditional cardiometabolic risk factors, limit precision in aging-related risk assessment. Here, we perform metabolomic analysis in 13,202 older adults from the natural aging cohort (NCT04517513). Leveraging a panel of 39 core metabolites, we develop accurate and interpretable machine learning models to identify metabolic dysfunction across different BMI categories, achieving area under the receiver operating characteristic curve (AUC) ranging from 0.763 to 0.801. These models reveal both shared metabolic alterations and obesity-specific changes (e.g., folate-mediated one-carbon metabolism). We further derive a BMI-metabolic health score (BMHS) that independently predicts all-cause and cardiovascular mortality beyond BMI-metabolic phenotypes, with improved stratification when combined with lifestyle factors. Our findings support a metabolomics-informed, behavior-modifiable strategy for precision prevention in aging populations, challenging BMI-centric paradigms and offering a scalable approach to evaluating metabolic health in late life.