AI model identifies lead contamination risk across cities with limited data

Lead sits in the pipes beneath American cities, invisible and patient. It leaches into drinking water through corroded service lines and old fixtures, and no one knows exactly where the problem is worst until someone tests the water—if they test it at all. In places like Flint, Michigan, where the water crisis became national news, and in smaller cities like Andover, Massachusetts, where contamination was discovered more quietly, public health officials face the same frustrating problem: they have spotty data from homeowners who bothered to sample their water, and no reliable way to predict which neighborhoods are at highest risk.

Researchers have now built a machine learning model that changes this equation. Instead of relying on scattered homeowner samples and traditional statistical tables that treat each property in isolation, they created a graph attention network—a type of artificial intelligence that understands how properties connect to one another through shared water infrastructure. The model learns by studying the spatial relationships between homes on the same block, the same street, the same water system. A contaminated pipe doesn't affect just one house; it affects everyone downstream.

The team trained their self-supervised graph attention network, or SSGAT, using data from nine different wards across Flint. They then tested whether what the model learned in Flint could work in Andover, a completely different city with its own water system, its own pipe network, its own contamination patterns. This is the hard test—can a model trained on one city's problem actually help another city?

It did. When applied to Andover without any additional training, the model achieved 81 percent accuracy in identifying which properties were at risk, with a macro-recall score of 0.66. That means it caught two-thirds of the contaminated properties, even when working with incomplete information. For comparison, the model performed nearly as well as a fully supervised approach that required much more labeled training data—the kind of data that most cash-strapped municipalities simply don't have. The spatial graph representation mattered. It outperformed traditional tabular models that ignored the fact that water systems are networks, not isolated points.

The researchers tested their predictions against multiple contamination thresholds, including the EPA's current action level of 10 parts per billion. At that threshold, the model's ability to flag at-risk properties held up. This matters because 10 ppb is the level at which the EPA says water systems must take action—must notify residents, must offer treatment, must investigate. But many cities don't know where to look first.

What makes this work significant is not just the accuracy number. It's the pathway it opens for cities with limited resources. Flint and Andover are not anomalies; they are examples of a much larger problem. Thousands of American municipalities have aging water infrastructure, incomplete testing data, and no systematic way to prioritize which neighborhoods need immediate attention. A model that can learn from one city's experience and transfer that knowledge to another, that can work with messy real-world data and still identify contamination hotspots with reasonable confidence—that is a tool that could actually be deployed.

The next question is whether it will be. The model works best when there is at least some baseline data to learn from, some testing history. In the most under-resourced communities, even that may not exist. But for the many cities caught between crisis and invisibility, between knowing there is a problem and knowing where it is, this approach offers a structured way forward. It suggests that the spatial logic of water systems themselves—the way pipes connect neighborhoods, the way contamination spreads—can be made legible to machines in ways that help humans see the risk more clearly.