Blood metal levels predict digestive tract cancer risk using machine learning in a U.S. cohort

Abstract

Background: Environmental metal exposure has been implicated in the development of digestive tract cancers, although the specific associations remain poorly defined. This study aimed to investigate the relationship between blood metal levels and the risk of digestive tract cancers among U.S. adults.

Methods: Data from the National Health and Nutrition Examination Survey (NHANES) 2011-2018, including 13,467 participants aged 20 years and older, were analyzed. Nine blood metals were measured. Multivariable logistic regression, restricted cubic spline models, and subgroup analyses were employed to assess the associations between metal levels and cancer risk. Additionally, a Random Forest (RF) model was used for cancer risk prediction.

Results: Among the participants, 9 had esophagus cancer (EC), 11 had gastric cancer (GC), and 83 had colorectal cancer (CRC). Compared to healthy controls, EC patients exhibited significantly higher blood levels of potassium (K, 4.40 vs. 4.00 mmol/L), cadmium (Cd, 12.46 vs. 2.49 µg/L), and lead (Pb, 0.09 vs. 0.05 µg/L). GC patients had elevated Pb levels (0.08 vs. 0.05 µg/L), while CRC patients showed higher concentrations of Cd (3.11 vs. 2.49 µg/L) and Pb (0.06 vs. 0.04 µg/L). Logistic regression analysis revealed significant associations between higher K (odds ratio [OR] = 7.58, 95% CI: 3.48-16.48, P < 0.0001), Cd (OR = 1.06, 95% CI: 1.04-1.08, P < 0.0001), and Pb (OR = 7.60, 95% CI: 3.26-17.72, P < 0.0001) levels and EC risk. Pb was also significantly associated with GC (OR = 5.26, 95% CI: 2.11-13.10, P < 0.001). The RF model showed an accuracy of 76% in predicting cancer risk, with SHapley Additive exPlanations (SHAP) analysis highlighting Cd and iron (Fe) as key contributors.

Conclusions: The study reveals a positive association between certain blood metals and digestive tract cancer risk, suggesting that limiting exposure to these metals may serve as a potential preventive measure.

Keywords: Blood metal; Colorectal Cancer; Esophagus cancer; Gastric cancer; Machine learning; Random Forest.

Fig. 1

Flowchart of study participants.

Fig. 2

Weighted values of blood metals for EC (A), GC (B) and CRC (C) in WQS models.

Fig. 3

ROC curves of K, Cd and Pd for the prediction of EC risk.

Fig. 4

Association of K (A), Cd (B) and Pb (C) with EC, Pb (D) with GC risk in RCS model.

Fig. 5

In-depth analysis of cancer-affected individuals based on machine learning models. (A) Spearman correlation matrix of input parameters; (B) optimization of hyperparameters for the Random Forest (RF) model based on accuracy; (C) confusion matrix of the RF model; (D) feature importance analysis based on SHAP.

Fig. 6

Partial dependence between different types of cancer and key features analyzed based on SHAP.