Baby Teeth Hold a Weekly Record of Lead Exposure, Revealing Lead Levels at Birth

Every baby tooth that a child loses is more than a keepsake. It is a biological archive, laid down layer by layer like the rings of a tree, recording the chemistry of the world the child lived in before and after birth. A new study published in Environmental Advances has now shown that this archive can be read with extraordinary precision: by measuring lead deposits in the dentine of shed baby teeth week by week, researchers were able to reconstruct, with the help of machine learning, what a child’s lead exposure was at the moment of birth — a level normally only obtainable from a blood sample taken from the umbilical cord.

The research, led by Marta Mainetti and Elena Colicino of the Icahn School of Medicine at Mount Sinai together with colleagues including Manish Arora and Robert O Wright, drew on the PROGRESS cohort, a prospective birth study in Mexico City that enrolled more than 1,000 pregnant women between 2007 and 2011. Lead remains a serious public health problem in Mexico: in 2023, an estimated 1.83 million Mexican children aged one to four — 17.4 percent of that age group — were affected by lead poisoning. Yet cord blood lead measurements, the standard biomarker of exposure around birth, are rarely collected in routine clinical practice anywhere in the world, leaving a critical gap in the exposure record during one of the most vulnerable windows of development.

The team’s approach exploits a quirk of basic chemistry. Lead mimics calcium in the body, slipping into calcifying tissues such as bone and tooth in place of the calcium that normally builds them. As teeth mineralize during the perinatal period, lead becomes locked into the dentine at the time of deposition. Using laser ablation-inductively coupled plasma mass spectrometry, or LA-ICP-MS, the researchers fired a nanosecond ultraviolet laser at thin sections of shed teeth, sampling fifty points of 35 micrometers each along the primary dentine near the enamel-dentine junction, and measured the normalized ratio of lead-208 to calcium-43 at each spot.

The real trick is dating. Teeth grow in recognizable increments: von Ebner’s lines form periodically during dentine deposition, much like tree rings, while a distinctive neonatal line — created by a disturbance in the secretory cells at the time of birth — cleanly divides the tissue mineralized before birth from that formed after. By counting these microscopic lines, the team assigned each laser measurement to a specific week, building a time series of lead exposure spanning from the twenty-first week before birth to the fifty-fifth week after. For the modeling, they focused on a 31-week window from 16 weeks before birth to 14 weeks after, and ultimately on a core 17-week window centered on delivery.

To translate those weekly dentine measurements into an estimate of cord blood lead, the researchers turned to machine learning. They compared several algorithms — Extreme Gradient Boosting (XGBoost), Bayesian Kernel Machine Regression, the Highly Adaptive Lasso, multivariate adaptive polynomial splines, generalized additive models, and linear regression — along with an ensemble SuperLearner that combined them all. The dataset comprised 258 mother-child pairs who had both weekly dentine lead data and a venous cord blood sample, analyzed on a log-2 scale to tame the skewed distribution of lead concentrations.

XGBoost emerged as the winner, not because it dramatically outperformed the ensemble but because it matched it while being simpler, more interpretable, and better at handling missing data. In 100 repeated training-validation splits, the model achieved a mean squared error of 0.554 on the log-2 scale, an area under the ROC curve of 0.842, and a mean absolute error of 12.9 micrograms per liter. The bias between predicted and observed cord blood values was essentially zero, the mean absolute percentage error was just 12.2 percent, and predicted and observed values correlated at 0.664 on the log-2 scale. When the model was fitted to the full dataset, performance improved further: the correlation rose to 0.780, the mean absolute error fell to 10.8 micrograms per liter, and the AUC reached 0.878.

Perhaps most striking was what the model learned about timing. Feature importance and SHAP analyses revealed that the second week before birth and the second week after birth were the most powerful predictors of cord blood lead concentration, with additional contributions from the sixth and eighth weeks before delivery and the week of birth itself. This finding dovetails with earlier work from the Mexican ELEMENT study, which found the strongest correlation between cord blood lead and dentine lead measured 12 to 15 days after birth. None of the demographic covariates — maternal age, socioeconomic status, second-hand smoke exposure, or fetal sex — contributed meaningfully to the predictions, underscoring that the signal comes from the tooth chemistry itself.

The clinical stakes are considerable. In the study population, the mean cord blood lead concentration was 33.5 micrograms per liter, and 32.6 percent of children exceeded the updated US CDC reference value of 35 micrograms per liter. The dentine-derived biomarker identified these high-exposure children with roughly 80 percent sensitivity and 73 to 80 percent specificity. Because lead exposure in fetal life and early childhood can permanently affect the nervous, immune, and cardiovascular systems, with effects that persist into adulthood, a tool that can retroactively flag children who experienced high exposure at birth — years after the fact, using a tooth the family kept in a drawer — could transform screening and prevention strategies, particularly for families living near smelters or with occupational take-home lead exposure.

Teeth offer practical advantages over blood as an exposure matrix. They are collected non-invasively at home, are less prone to contamination, and preserve their chemical record for years, whereas lead persists in blood for only about 35 days. The authors caution that their model was trained on a Hispanic population with moderate exposure levels, that the sample size of 258 is modest, and that missing dentine data increased with distance from birth because different tooth types mineralize at different stages of pregnancy. Sensitivity analyses — including models run without imputed values and with a wider 31-week exposure window — confirmed the main findings, and the researchers note that future studies with larger cohorts could adopt more robust validation splits.

The study provides a proof of concept that weekly lead measurements locked in baby teeth can stand in for a cord blood sample that was never taken. As the authors conclude, this dentine-derived biomarker may allow clinicians and researchers to identify children at risk of lead-induced conditions long after birth, opening a path toward regulatory standards and early interventions for a toxic exposure window that has, until now, remained largely invisible.

Subject of Research: Estimating cord blood lead concentrations from weekly micro-spatial dentine lead measurements in deciduous teeth using machine learning

Article Title: Dentine-derived lead exposure biomarker at birth: an estimation of cord blood concentrations through micro-spatial weekly child dentine lead measures

Article References: Mainetti, M., Saddiki, H., India-Aldana, S., Tellez-Rojo, M. M., Olascoaga, L. T., Wright, R. O., Arora, M., & Colicino, E. (2026). Dentine-derived lead exposure biomarker at birth: an estimation of cord blood concentrations through micro-spatial weekly child dentine lead measures. Environmental Advances, Article 100761. https://doi.org/10.1016/j.envadv.2026.100761

Image Credits: AI Generated

DOI: 10.1016/j.envadv.2026.100761

Keywords: lead exposure, baby teeth, dentine, cord blood, machine learning, XGBoost, PROGRESS cohort, LA-ICP-MS, biomarker, prenatal exposure, environmental health, Mexico