Explainable machine learning for identifying osteoarthritis comorbidity among adults with diabetes: A nationally representative NHANES study

Objectives Diabetes and osteoarthritis (OA) frequently co-occur, yet scalable approaches for identifying OA comorbidity among adults with diabetes remain limited. We developed and interpreted machine-learning models to support cross-sectional identification of OA comorbidity among adults with diabetes. Methods We analyzed seven cycles of the National Health and Nutrition Examination Survey (NHANES) (2005–2018). Features were preprocessed using variance inflation factor screening, Synthetic Minority Oversampling Technique (SMOTE), and Boruta feature selection. Weighted quantile sum (WQS) regression assessed nutrient-mixture associations. Six machine-learning models were trained under two feature settings: combined demographic and dietary features and dietary-only features. Model performance was evaluated using AUC-ROC, PR-AUC, accuracy, F1 score, sensitivity, and specificity. SHAP and LIME were applied for explainable model interpretation. Results Nutrient mixtures showed inverse but non-significant associations with diabetes–OA comorbidity in WQS regression (unconstrained OR = 0.86, 95% CI 0.73–1.01; constrained OR = 0.98, 95% CI 0.89–1.07). Random Forest performed best with combined features, whereas XGBoost performed best with dietary-only features. Explainability analyses consistently highlighted magnesium and caffeine as influential predictors in the combined model, while theobromine, niacin, and alpha-tocopherol were dominant drivers in the dietary-only model. Conclusion Explainable machine learning using dietary and demographic variables may facilitate identification of OA comorbidity among adults with diabetes. The identified nutrient signals warrant further validation and may inform future decision-support and personalized nutrition research.