355 - A Machine-Learning Model Predicts Census-Tract Risk for Pediatric Asthma Emergency Department Visits Using Place-Based Social Determinants of Health, Community Health Indicators and Clinical Variables

Poster Number: 355
Publication Number: 355.302

Yolande M. Pengetnze, PCCI, Flower Mound, TX, United States; Thomas E. Roderick, Parkland Center for Clinical Innovation, Frisco, TX, United States; Hannah Ligon, PCCI, Dallas, TX, United States; Yusuf Tamer, PCCI, Dallas, TX, United States

Background: Asthma affects ~6 Million US children, with a disproportionate impact on minority children from low socioeconomic background. Place-based social determinants of health (SDoH) and community health status are known correlates of asthma health disparities. No published study, however, has leveraged SDoH, community and clinical indicators to predict pediatric asthma emergency department visits (EDV) at the community level.

Objective: To develop a model incorporating place-based SDoH, community indicators, and clinical variables to predict census-tract risk for pediatric asthma EDV/hospitalizations

Design/Methods: A retrospective population analysis across 527 Dallas County census tracts, 01/2018 to 10/2021.
Data Sources:
Dallas Fort Worth Hospital Council - North Texas HIE (health services utilization)
Parkland Health - Dallas County safety net health system (clinical indicators)
2019 American Community Survey (socioeconomic data)
CDC's Housing & Transportation + PLACES (community health status).
NASA (air quality)

Outcome: number of census-tract children with asthma-related EDV or hospitalizations within 90 days per 10,000 population.
Predictors: census-tract health services utilization, place-based SDoH (e.g., housing/transportation), chronic/mental health conditions prevalence, air quality, medication use;  
Confounders: Seasons and measurement year.
Data was aggregated week-over-week over 183 weeks and split 80/20 into Training & Test sets. Bivariate analyses in Training set identified candidate predictors, which were fed into LASSO regression for multivariable model selection. The model was run weekly and validated using Test set adjusted R-Squared.

Results: Of 2.6 million Dallas County residents, 30% are children ( >780,000), 40% Latino and 23% Black. 60,796 pediatric asthma EDV and 8,989 hospitalizations were observed. Bivariate analyses identified 66 (of 184) candidate predictors for LASSO selection of which ten were retained in the final model. Household access to automobile, controller medication use, and female gender were protective (Coefficient [-0.05], [-0.06], and [-0.03], respectively; p< .01). Recent 90-day EDV, COPD, PM 2.5, Black & Latino were risk factors (Coeff. [0.15], [0.17], [0.01], [0.23], and [0.05], respectively; p< .01). The model explained 56% of outcome variability (Adjusted R-Squared = 0.56, p< .001).

Conclusion(s): A machine-learning model combining place-based SDoH, clinical and community health indicators accurately predicts census-tract risk for pediatric asthma EDV/hospitalizations, providing actionable and scalable insights to support equitable public health prioritization.