Self-supervised graph attention networks for community-engaged lead contamination risk assessment

Abstract

Lead contamination in residential water infrastructure poses a persistent public health risk, yet current assessment approaches rely on sparse homeowner sampling and tabular models that do not explicitly capture spatial dependencies in shared urban systems. Here we present a self-supervised graph attention network (SSGAT) that models lead contamination risk using property-level graph structure and reconstruction-based pretraining. Using leave-one-ward-out validation across nine wards in Flint, Michigan, followed by cross-city transfer to Andover, Massachusetts, SSGAT achieved a macro-recall of 0.66 ± 0.03 and an accuracy of 0.81± 0.02 in the transfer setting. Performance was statistically comparable to a fully supervised graph model (macro-recall 0.69 ± 0.04, p = 0.30) and improved contaminated-property detection relative to tabular baselines. Results were evaluated under multiple concentration thresholds, including the current EPA action level of 10 ppb (Parts per billion). These findings indicate that spatial graph representations can support cross-city generalization under class-imbalanced conditions, offering a structured approach for contamination screening in data-limited jurisdictions.