Precision occupational lead exposure assessment through medical-informed machine learning

Precise assessment of occupational lead exposure remains a major global public health challenge, as existing assessment methods are limited by overreliance on single biomarkers, insufficient capture of complex multi-factor interactions, and algorithmic bias driven by severe class imbalance. To address these gaps, this study developed a novel medical prior-informed ACO-RF model with a dual-layer optimization system anchored in lead's ADME toxicokinetic pathways. As the first ACO-RF framework tailored for this field, it encodes lead's Absorption-Distribution-Metabolism-Excretion (ADME) pathways into algorithm design, unlike engineering-focused variants. When validated on a multicenter cohort of 2867 lead-exposed workers from 21 enterprises in Jiangsu Province, China, ACO-RF reduced feature dimensionality by over 75% from 79 to 19 for blood lead (BL) prediction and to 16 for urine lead (UL) prediction, while achieving superior predictive performance with AUC-ROC values of 0.9646 for BL and 0.9319 for UL. Critically, the optimized feature subset is highly aligned with lead toxicokinetic mechanisms, ensuring strong biological interpretability, and the model significantly improved the recall of abnormal BL samples from 0.409 to 0.841, markedly reducing missed diagnoses of subclinical exposure. This work provides a robust, interpretable tool for early lead exposure risk stratification, and establishes a generalizable knowledge-guided machine learning framework for environmental and occupational health research.