Associations between exposure to heavy metal and sarcopenia prevalence: a cross-sectional study using NHANES data

Abstract

Background: Sarcopenia is a condition that adversely affects individuals' quality of life and physical health. Exposure to heavy metals poses a significant risk to human health; however, the impact of heavy metal exposure on sarcopenia remains unclear. Therefore, this study expects to construct a risk prediction machine model of heavy metal exposure on sarcopenia and to interpret and analyze it.

Methods: Model construction was based on data from the NHANES database, covering the years 2011 to 2018. The predictor variables included BA, CD, CO, CS, MN, MO, PB, SB, SN, TL, and W. Additionally, demographic characteristics and health factors were included in the study as confounders. After identifying the core variables, optimal machine learning models were constructed, and SHAP analyses were performed.

Results: We found that the LGBM model exhibited the best predictive performance. SHAP analysis revealed that TL, SN, and CS negatively influenced the prediction of sarcopenia, while CD positively contributed to it. Additionally, le8 BMI was the covariate that had the most significant positive impact on the prediction of sarcopenia in our model.

Conclusion: For the first time, we have developed a machine learning (ML) model to predict sarcopenia based on indicators of heavy metal exposure. This model has accurately identified a series of key factors that are strongly associated with sarcopenia induced by heavy metal exposure. We can now identify individuals at an early stage who are suffering from sarcopenia due to heavy metal exposure, thereby reducing the physical and economic burden on public health.

Keywords: NHANES; SHAP; heavy metal exposure; machine learning; sarcopenia.

Figure 1

Flowchart of the study design.

Figure 2

Forest plot of logistic regression results showing odds ratios for sarcopenia across quartiles of metal exposure.

Figure 3

Boruta algorithm feature screening plot. (A) Boruta feature screening importance score plot. (B) Corresponding importance box plots for each variable of Boruta feature screening.

Figure 4

Lasso algorithm eigenvalue screening plot. (A) Lasso regression Lambda and CVM plot. (B) Lasso regression Lambda and coefficients plot.

Figure 5

Boruta and Lasso feature selection overlap.

Figure 6

ROC curves of the train and test sets of 6 ML models. (A) ROC curves of the train set. (B) ROC curves of the test set.

Figure 7

ROC curve for the 5-fold test.

Figure 8

ROC curve for Bootstrap evaluation.

Figure 9

SHAP diagram. (A) SHAP importance plot; (B) SHAP bees plot; (C) SHAP heatmap; (D) SHAP line plot; (E) SHAP heat plot with age, race, Le8 BMI, Le8 smoke, Le8 pa, CD, CS, SN, and TL.

Figure 10

Logistic regression forest plot based on gender differences.

Figure 11

Logistic regression forest plot based on age differences.

Figure 12

Logistic regression forest plots based on racial differences.

Figure 13

Logistic regression forest plot based on dietary differences.

Figure 14

Logistic regression forest plots based on differences in the prevalence of chronic diseases.