Answers are grounded in this paper and cite their sources.
Environmental chemical exposures may contribute to cancer-related outcomes, yet population-based integration of high-dimensional biomonitoring data with interpretable machine learning remains limited. Using NHANES 2011–2012, we analyzed 1,443 adults and 70 biospecimen-measured environmental chemicals across 10 exposure classes in relation to prevalent self-reported cancer history. Eight algorithms—support vector machine (SVM), k-nearest neighbors (KNN), random forest (RF), neural network (NNET), logistic regression (LR), naïve Bayes (NB), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost)—were developed within the training set using repeated nested cross-validation. The best-performing tree-based model in training-side validation was selected for SHapley Additive exPlanations (SHAP). After model specifications were frozen, the held-out testing set was used for final performance evaluation and out-of-sample SHAP interpretation, but not model development. SHAP-prioritized chemicals were assessed using exploratory survey-weighted logistic regression and restricted cubic splines. LightGBM met the prespecified tree-model selection criterion and achieved the highest held-out ROC-AUC (0.838), followed by XGBoost (0.812). Among the top 15 SHAP-ranked predictors, 10 were environmental chemicals. Urinary lead (OR = 1.41, 95% CI 1.19–1.67), DHBMA (1.60, 1.09–2.35), CEMA (1.52, 1.08–2.12), and urinary tin (1.45, 1.13–1.87) showed positive per-1-SD associations that remained significant after Benjamini–Hochberg false discovery rate correction and were further supported by sensitivity analyses. Spline analyses showed overall associations for these biomarkers, with CEMA exhibiting nonlinearity. This framework supports exploratory exposure prioritization and hypothesis generation. Prospective studies with repeated exposure assessment and validated cancer diagnoses are warranted.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: