Composite dietary antioxidant index and HPV infection from single and mixed associations to SHAP-interpreted machine learning predictions

Abstract

Background: Some studies have shown that dietary antioxidants may prevent the occurrence of Human Papillomavirus (HPV) infection. However, the relationship between the composite dietary antioxidant index (CDAI) and HPV infection among adult women in the United States remains unknown.

Methods: Participants from the National Health and Nutrition Examination Survey (NHANES) during 2003-2016 were included. Multivariable logistic regression, restricted cubic spline (RCS) regression, weighted quantile sum (WQS) regression, and Bayesian kernel machine regression (BKMR) were used to analyze the associations between CDAI and its sub-components and HPV infection. In addition, nine machine learning (ML) methods were employed to construct predictive models, and SHapley Additive exPlanations (SHAP) was used to further interpret the optimal model.

Results: This study enrolled 9,224 adult female participants. After adjusting for multiple confounding variables, CDAI was independently negatively associated with HPV infection (OR: 0.98, 95%CI: 0.97-0.99, p = 0.01). RCS indicated an L-shaped association between CDAI and HPV infection. In the WQS model, the WQS index of CDAI was still robustly negatively associated with HPV infection (OR: 0.78, 95%CI: 0.71-0.86, p < 0.0001). In the mixture effect, BKMR analysis confirmed the negative association between six antioxidants and HPV infection. Both WQS and BKMR confirmed that vitamin E had the strongest negative association with HPV infection. Additionally, among the nine machine-learning models, the Gradient Boosting Machine (GBM) showed the best predictive performance [area under curve (AUC) = 0.685]. SHAP analysis indicated that marital status, smoking, drinking, race, age, and CDAI had a significant impact on the model's prediction.

Conclusion: Antioxidant-rich diets, especially increased intake of vitamin E, are significantly negatively associated with HPV infection. A GBM model with 12 features can effectively predict the occurrence of HPV infection, among which CDAI is an important factor in the model.

Keywords: BKMR; HPV infection; NHANES; SHAP; WQS; composite dietary antioxidant index; machine learning.

Figure 1

Dose–response relationship between CDAI and HPV infection. Model were adjusted for age, race, education attainment, marital status, poverty-income ratio, smoking, drinking status, BMI, energy intake, physical activity, hyperlipidemia, hypertension, DM, cancer, CVD, CKD, liver problem. CDAI, composite dietary antioxidant index; HPV, human papillomavirus; BMI, body mass index; CVD, cardiovascular disease; CKD, chronic kidney disease; DM, diabetes mellitus; OR, odds ratio; CI, confidence interval.

Figure 2

Single and mixed association between CDAI and HPV infection in BKMR and WQS model. (A) Weights of WQS index of six dietary antioxidants in HPV infection in WQS model. The dashed black lines represent the cutoff to discriminate which element has a significant weight. (B) Mixed association between CDAI and HPV infection in BKMR model. (C) Single-exposure effects of individual dietary antioxidants on HPV infection in BKMR model. (D) Single exposure dose relationship of individual dietary antioxidants with HPV infection in BKMR model. All models were adjusted for age, race, education attainment, marital status, poverty-income ratio, smoking, drinking status, BMI, energy intake, physical activity, hyperlipidemia, hypertension, DM, cancer, CVD, CKD, liver problem. CDAI, composite dietary antioxidant index; HPV, human papillomavirus; WQS, weighted quantile sum; BKMR, Bayesian kernel machine regression; BMI, body mass index; CVD, cardiovascular disease; CKD, chronic kidney disease; DM, diabetes mellitus.

Figure 3

Boruta algorithm for selecting ML model variables. ML, machine learning; BMI, body mass index; CVD, cardiovascular disease; CKD, chronic kidney disease; DM, diabetes mellitus; PIR, poverty-income ratio; CDAI, composite dietary antioxidant index.

Figure 4

ROC curves in the training set and testing set. (A) ROC curves of the training set. (B) ROC curves of the testing set. ROC, receiver operating characteristic curve; AUC, Area Under Curve.

Figure 5

SHAP diagram for interprete GBM model. (A) SHAP honeycomb diagram of the GBM model. (B) SHAP value ranking of the variables in the model. (C) SHAP force plot for the twelfth sample in the study population. SHAP, SHapley Additive exPlanations; GBM, Gradient Boosting Machine; CDAI, composite dietary antioxidant index; DM, diabetes mellitus; CVD, cardiovascular disease; PIR, poverty-income ratio.