Interpretable machine learning approaches for predicting prostate cancer by using multiple heavy metal exposures based on the data from NHANES 2003–2018

Environmental pollution plays a major role in the development of prostate cancer (PCA). However, there has been no research on machine learning (ML) modelling between multiple heavy metal exposures and PCA risk. Based on the 8022 samples from the 2003-2018 National Health and Nutrition Examination Survey (NHANES) database, we utilized the information pertaining to the concentrations of 18 blood and urinary heavy metals and minerals as well as 14 covariates. Among the eight ML models evaluated, the random forest (RF) algorithm showed superior performance, achieving an accuracy of 72.835 %, an area under the receiver operating characteristic curve (AUC) of 0.869, an F₁ score of 0.145, a G-mean of 0.749, and a Youden index of 0.498 in the test set. Four interpretable methods were integrated into the ML model. RF found that specific levels of blood lead (Pb) (0.449-29.964 µg/dL), urinary cesium (Cs) (1.822-270.426 µg/L), and urinary antimony (Sb) (0.015-4.953 µg/L) were positively associated with the PCA risk, while blood cadmium (Cd) (0.247-9.025 µg/L) showed a negative association. Notably, urinary Cs and Sb emerged as novel risk-related metals for the PCA in our study. The synergistic effect analysis further identified blood Pb, urinary Sb, and urinary Cs as the major contributing factors. The predictive model established in this study can provide valuable strategies for the prevention and the control of PCA.