Skip to main content
Evidence-Based Supplement Research
Evidence-Based Supplement Research
Pillser
Search
Sign UpLog In

Dietary antioxidants and CKM-depression comorbidity: a primary analysis with a secondary evaluation of all-cause mortality using six machine learning algorithms.

  • 2025-11-27
  • European journal of medical research 30(1)
    • Yi Tang
    • Zilong Yue
    • Juncang Wu
    • Hongwei Liu
    • Qiuwan Liu
    • Kangrui Zhang
    • Yueyu Zhang
    • Yu Wang
    • Shuaizhou Wang
    • Xinyi Chen
    • Xun He
    • Jie Hu
    • Zhinan Ye

Study Design

Type
Observational
Population
US adults
Methods
Observational analysis of NHANES 2007-2010 with linked mortality; LASSO and Boruta feature selection, six classifiers, multivariate logistic regression, Cox proportional hazards models
Funding
Unclear

Background

We examined whether dietary antioxidant intake is associated with CKM-depression comorbidity (primary outcome) and all-cause mortality (secondary outcome) in US adults.

Methods

In an observational analysis of NHANES 2007-2010 with linked mortality, we defined the primary outcome as CKM-depression comorbidity and the secondary outcome as all-cause mortality. Predictors were selected using LASSO (linear models) and Boruta (tree-based models). Six classifiers (AdaBoost, GBDT, LGBM, RF, SVM, NB) were trained. Evaluation used stratified tenfold/nested cross-validation; all preprocessing, feature selection, and SMOTE were performed inside training folds only to avoid leakage. Discrimination metrics included ROC-AUC, PR-AUC, F1, MCC, recall, specificity; calibration was assessed on out-of-fold probabilities using Brier score and calibration slope/intercept, with reliability curves. Subsequently, associations were assessed using multivariate logistic regression (for comorbidities) and Cox proportional hazards models (for all-cause mortality). We used permutation-adjusted P-values for multiplicity, restricted cubic splines for dose-response, and Schoenfeld residuals for the proportional-hazards assumption. E-values quantified residual confounding.

Results

Among ML models, LGBM performed best (test AUC = 0.978) with favorable calibration. In fully adjusted models, higher anthocyanidin intake was associated with lower odds of CKM-depression comorbidity: peonidin (OR = 0.716, 95% CI 0.572, 0.897), petunidin (OR = 0.749, 95% CI 0.633, 0.885), and total anthocyanidins (OR = 0.973, 95% CI 0.959, 0.988). For all-cause mortality, Cox models showed inverse associations for petunidin (HR = 0.936, 95% CI 0.900, 0.975) and total anthocyanidins (HR = 0.993, 95% CI 0.988, 0.998), whereas peonidin was not significant after permutation adjustment (HR = 0.964, 95% CI 0.929, 1.000). Restricted cubic splines suggested monotonic inverse trends; proportional-hazards tests showed no violations.

Conclusions

Specific anthocyanidins were cross-sectionally associated with lower CKM-depression comorbidity and, for petunidin and total anthocyanidins, with lower mortality hazard. Given the observational design and potential residual confounding, these findings should be interpreted as associations rather than causal or protective effects.

Research Insights

Back to top