LocaliType: Development of locality-based social determinants of health archetypes using census tract data from the...

Locality archetypes (denoted as LocaliTypes) - neighborhood, district, or nearby region patterns shaped by multiple social determinants of health (SDOH) - offer a novel approach to understanding place-based inequalities while preserving data confidentiality in the era of big data. Using 2020 U.S. Census shaped census tract-level data from the UCSF Health Atlas, we developed LocaliTypes and conducted exploratory validation analyses. We grouped 107 variables spanning race/ethnicity, socioeconomic conditions, environmental exposures, and disaster risk into five domains using large language model-based text-embeddings. Within each domain, dimensionality was reduced using principal component analysis, followed by unsupervised deep learning models to identify LocaliTypes. To assess validity, we examined differences in historical redlining, 15 health outcomes, and established neighborhood indices across LocaliTypes, with adjustment for multiple comparisons. The dataset included 77,054 census tracts representing 92.6% of the US population and revealed high uneven distributions of key SDOH indicators (e.g., poverty among all families: median 6.8%, range 0% - 91.5%). We identified 11 LocaliTypes with randomly assigned labels that exhibited marked heterogeneity in deprivation, health outcomes, and historical disadvantage (global testing p < 0.001). Pairwise comparisons revealed substantial heterogeneity across LocaliTypes in both opportunities and deprivation, as reflected in existing indices. Also, more deprived LocaliTypes had substantially higher percentages of historically redlined neighborhoods (grade D) and a greater burden of chronic health conditions. By capturing the multidimensional and spatially structured nature of SDOH, these locality archetypes provide a scalable framework for studying place-based health inequities and supporting policy-relevant research using de-identified data.