Precision risk assessment for pediatric hospitalization using address-level data in Cincinnati, Ohio

This study leveraged residence- and neighborhood-specific socio-environmental data linked to population-wide healthcare data to characterize risk for pediatric hospitalization for every residential address in Cincinnati, Ohio.

We linked hospitalization data (07/01/2016-06/30/2022) to parcel-level housing data from the Hamilton County Auditor and Cincinnati Department of Buildings & Inspections and block-level crime data from the Cincinnati Police Department. Addresses were localized to 2010 census tracts to join variables from the U.S. Census American Community Survey and Eviction Lab. Generalized random forest models predicted hospitalization risk and birth-adjusted hospitalization risk scores, accounting for child residency using vital birth records. Model performance was assessed based on predictions calibrated using varying diagnostic thresholds; fairness was evaluated based on block-level racial demographics.

We matched 81.5% of hospitalizations to residential addresses. Among 77,077 total addresses, 7.4% had ≥1 hospitalization. Our model performed well (AUC: 0.98-0.99) in characterizing high-risk addresses, with housing violations, violent crime, and market total value among top predictors. The birth-adjusted model also showed high performance (AUC: 0.92-0.93) and moderate agreement with the hospitalization risk model (κ=0.43), emphasizing different address types.

The strong performance of both models highlights the potential of address-level predictions and multiscale data integration to build on traditional area-level analyses and advance precision population health. Future directions include geographic expansion, incorporation of additional predictors and outcomes, stakeholder engagement, and patient-level validation.

This work offers a scalable, privacy-preserving approach to identifying pediatric health risks with high spatial precision, supporting targeted clinical and policy interventions.