A Metabolomics-driven machine learning model for osteoporosis risk prediction

Abstract Osteoporosis, marked by decreased bone mineral density (BMD), poses a major public health concern by increasing fracture risk, lowering quality of life, and raising healthcare costs in aging populations. Accurate risk prediction is essential for early diagnosis and targeted intervention. While machine learning (ML) models have been used to predict osteoporosis risk based on clinical factors, their accuracy and generalizability are limited. Multi-omics studies, especially in genomics and proteomics, have shown promise in predicting osteoporosis-related traits such as BMD and fracture risk. However, the integration of metabolomics, critical for bone metabolism, remodeling, and mineralization remains underexplored in this context. This study aimed to identify novel metabolites associated with osteoporosis risk and assess their predictive utility. Using hip BMD measurements, clinical data, and metabolomic profiles from 2,041 participants aged 40 and older, we developed a predictive model integrating metabolomics with clinical risk factors. Model performance was evaluated using the area under the receiver operating curve (AUC). We identified 44 metabolites significantly associated with osteoporosis risk (p 0.05), many enriched in small molecule transport and metabolic pathways. Notably, 25 had prior links to bone metabolism. Incorporating these metabolites into our metabolomics-driven ML model improved predictive performance compared to a clinical-only model (AUC clinical + metabolomics = 0.763 vs. AUC clinical = 0.741, p = 0.017). Feature importance analysis identified five key clinical predictors (weight, grip strength, height, age, and sex) and five top metabolites (N-acetylcarnosine, hypotaurine, homoarginine, epiandrosterone sulfate, and dehydroepiandrosterone sulfate). Mendelian randomization confirmed causal effects of these top metabolites on BMD variation. We identified a metabolomic signature for osteoporosis risk and demonstrated that integrating metabolomics with clinical data significantly enhances ML-based prediction. These findings point to the promise of metabolomics in advancing strategies for early detection and personalized care in aging populations at risk for osteoporosis.