Interpretable machine learning in type-2 diabetes prediction in patients with depressive symptoms: Insights from mendelian randomization and physical activity

Objective This study aimed to examine the causal relationship between depression and type-2 diabetes (T2D) and to develop a machine-learning (ML)-driven risk prediction model for T2D. Methods This cross-sectional study performed forward and reverse Mendelian randomization (MR) and multivariable MR (MVMR) analyses using GWAS summary data. Data from the National Health and Nutrition Examination Survey 2007-2018 were used to develop a ML–based T2D risk prediction model among individuals with depressive symptoms ( n = 790). Nine ML algorithms were compared, incorporating Boruta feature selection, 10-fold cross-validation, and SHapley Additive ExPlanations (SHAP) analysis to identify and interpret key predictors. Moderation analysis was conducted to evaluate whether the triglyceride-glucose (TyG) index and physical activity (PA) influence the relationship between the predictor and the outcome. Results Inverse variance weighted MR analysis showed that genetic liability to depression increased the risk for T2D (OR = 1.38, 95% CI: 1.01–1.90), while no significant causal effect was found in the reverse direction. MVMR analysis showed that PA was inversely related to T2D risk (OR = 0.98, 95% CI: 0.96–0.99, p = 0.001). Boruta feature selection identified age and the TyG index as significant predictors. Among nine ML algorithms, LightGBM achieved the most consistent performance, with AUROC values of 0.84 (training) and 0.75 (testing), and PRAUC values of 0.977 and 0.956, respectively. SHAP analysis confirmed that the TyG index had the highest feature importance (mean SHAP value = 0.099), followed by age (0.025). Moderation analysis identified a significant interaction between the TyG index and PA (ΔR 2 = -0.007, F = 4.89, p = 0.027). Conclusions Our MR results support a causal effect of depression on T2D risk, with the TyG index and physical activity (PA) serving as modulators. Our ML–driven prediction model provides an interpretable screening tool for T2D risk in patients with depressive symptoms.