Multi-media biomarkers

Abstract

OPS 03: Machine learning in environmental epidemiology, Room 315, Floor 3, August 26, 2019, 4:30 PM - 5:30 PM Background/Aim: Exposure assessment traditionally relies on measured chemical concentrations in a single biological medium. Toxicokinetic properties cause chemicals to distribute unevenly among different media, thus each medium provides incomplete information about body burden. We propose that machine learning and statistical approaches developed for mixtures can create aggregated exposure estimates, which we term multi-media biomarkers (MMBs), for a given chemical that outperform exposure estimates from individual media. Methods: We measured lead (Pb) in blood, hair, nails, and urine to derive aggregated MMBs and then tested their association with Wechsler Intelligence Scale for Children (WISC) total IQ scores in the Public Health Impact of Metals Exposure (PHIME) project, a study of Italian adolescents aged 10-14 years (N~720). We used a supervised learning technique, weighted quantile sum regression (WQS), and an unsupervised learning technique, independent component analysis (ICA) to create the Pb MMB. Since the MMB produced using ICA is not normally distributed, we measured adjusted associations between Pb and IQ using Spearman correlation and residuals generated from a model where sex and age are predictors of IQ. Results: Total IQ was associated with the Pb MMB for ICA (r=-0.14; p=0.03) and for WQS (r=-0.16; p=0.01). Total IQ was not associated with Pb measured using the individual exposure biomarkers (blood: r=-0.10; p=0.11, urine: r=0.09; p=0.13, hair: r=-0.10; p=0.11, nails: r=-0.06; p=0.35). Weights highlighting the contributions to each MMB indicate that nails (59%) and hair (23%) contributed most for WQS, whereas blood (51%) and nails (21%) contributed most for ICA. Conclusions: These results indicate that MMBs can be more effective measures of exposure than individual biomarkers. Though MMBs have increased assay costs, they may increase statistical power when the choice of exposure medium is unclear. Future work will need to validate these methods for other toxicants and evaluate the associations with WISC subscales.