'Pb-predict’: using machine learning to locate lead plumbing in a large public water system

[摘要] Struggling to respond to elevated lead levels in residential tap water, cities like Flint, MI and Pittsburgh, PA are undergoing large-scale efforts to remove the lead pipes that bring water service to their customers. However, limited geographic data on plumbing materials throughout housing stocks represents a logistical challenge for local authorities to locate and replace lead service lines. This study tests whether available geographic data on housing conditions and plumbing materials can effectively inform risk assessment and thus, expedite replacement programs and help prevent exposure to lead. To do so, we train and compare multiple types of machine learning classification algorithms to predict the presence or absence of lead service lines at properties in Pittsburgh. The results show that the probability of having a lead service line increases for houses built before 1930 and demonstrate the significance of parcel age, spatial proximity and other housing characteristics as predictive features for locating lead in water hazards. Accurate targeting of high-risk housing units may inform the strategy of decision-makers working to ensure that residents of aging American homes have safe drinking water. Therefore, the results are mapped to simulate the prevalence of lead service lines throughout the City of Pittsburgh and a framework for other cities is discussed.

[发布机构] the University of Pittsburgh