Childhood lead exposure affects over 500,000 children under 6 years old in the US; however, only 14 states recommend regular universal blood screening. Several studies have reported on the use of predictive models to estimate lead exposure of individual children, albeit with limited success: lead exposure can vary greatly among individuals, individual data is not easily accessible, and models trained in one location do not always perform well in another. We report on a novel approach that uses machine learning to accurately predict elevated Blood Lead Levels (BLLs) in large groups of children, using aggregated data. To that end, we used publicly available zip code and city/town BLL data from the states of New York (n = 1642, excluding New York City) and Massachusetts (n = 352), respectively. Five machine learning models were used to predict childhood lead exposure by using socioeconomic, housing, and water quality predictive features. The best-performing model was a Random Forest, with a 10-fold cross validation ROC AUC score of 0.91 and 0.85 for the Massachusetts and New York datasets, respectively. The model was then tested with New York City data and the results compared to measured BLLs at a borough level. The model yielded predictions in excellent agreement with measured data: at a city level it predicted elevated BLL rates of 1.72% for the children in New York City, which is close to the measured value of 1.73%. Predictive models, such as the one presented here, have the potential to help identify geographical hotspots with significantly large occurrence of elevated lead blood levels in children so that limited resources may be deployed to those who are most at risk.