Machine learning-based prediction of osteoporosis in prefrail and frail older adults: insights from SHAP analysis of dietary antioxidants

Osteoporosis poses significant risks in frail older adults. While oxidative stress influences bone metabolism and dietary antioxidants may provide protection, their predictive value for osteoporosis risk remains underexplored. To develop machine learning models predicting osteoporosis in frail older adults and evaluate dietary antioxidant contributions using SHapley Additive exPlanations (SHAP). Data from NHANES 2007–2010 and 2017–2018 cycles included 29,940 initial participants. After excluding those aged <65 years, missing bone mineral density/osteoporosis data, and non-frail individuals, 2,160 frail older adults remained. Further excluding incomplete dietary data yielded 1,829 participants. Sixty-six candidate features – demographic, clinical, and dietary antioxidant variables – were assessed. Highly correlated features (r > 0.9) were removed, class imbalance addressed using SMOTE, and seven machine learning algorithms compared. SHAP analysis interpreted the best-performing model. Random forest achieved optimal performance (accuracy = 0.9374; AUC = 0.9631; sensitivity = 0.8919). SHAP analysis identified dietary antioxidants – selenium, zinc, vitamin E as key predictors alongside age and gender. Lower antioxidant intake correlated with higher osteoporosis risk. Cross-validation confirmed model robustness. In this NHANES-based dataset, dietary antioxidants emerged as important predictors within machine learning models for osteoporosis risk in prefrail and frail older adults. Random forest combined with SHAP interpretation provided an interpretable framework for exploratory risk stratification, although external validation is required before clinical application.