Refining Small-Area Cancer Surveillance Through AI-Driven Spatial Clustering - PROJECT SUMMARY County boundaries are the dominant unit for cancer surveillance in the United States, yet they are poorly suited for capturing meaningful geographic variation in population cancer burden. In states like Texas, home to both large, demographically complex urban centers and sparsely populated rural areas, county-level reporting often obscures critical geographic differences. A large proportion of rural counties frequently face data suppression due to small populations and incident cases of cancers, while large urban counties mask high-burden neighborhoods behind aggregated statistics. These limitations hinder public health efforts to identify, monitor, and respond to geographic differences in cancer outcomes. This is especially problematic for cancer prevention and control efforts of breast, colorectal, and lung cancers, which account for nearly 50% of all cancer incidence and 45% of deaths annually in Texas. Previous studies have explored alternatives to county- based cancer reporting by using spatial clustering methods that aggregate census tracts or ZIP codes. However, these approaches typically rely on static, user-defined rules and thresholds, limiting their ability to balance critical trade-offs—such as minimizing data suppression, maximizing geographic granularity, and ensuring demographic homogeneity needed for responsive public health action. To address this critical gap, we propose a novel Geographic Artificial Intelligence (GeoAI) approach, specifically the application of Neural Cellular Automata and Graph Neural Networks, to create sub-county geographic boundaries to improve cancer surveillance using individual-level cancer incidence data from the Texas Cancer Registry (2018–2022). The first aim of our project will develop and test GeoAI-based sub-county geographic zones for breast, colorectal, and lung cancer surveillance that minimize data suppression, improve demographic homogeneity, and enhance spatial coherence across Texas. In our second aim, we will leverage the newly created GeoAI- defined zones to examine geographic variation in late-stage incidence and cancer-specific mortality, identifying high-burden areas obscured by traditional county boundary reporting. This project represents the first use of Neural Cellular Automata and Graph Neural Networks to construct flexible, data-driven geographic units optimized for cancer surveillance. These methods will allow us to generate zones that are population-stable, demographically meaningful, and tailored to local context. At the end of the study, we will have created a generalizable framework that reduces data suppression, uncovers hidden geographic cancer variation, and improves the spatial precision of cancer surveillance, offering new tools for targeting prevention and control efforts at the community level.
