Construction of a Pediatrics Risk Score to Predict High Health Care Costs Among a Community Health Center Cohort

Abstract

Risk-stratification strategies are needed for ambulatory pediatric populations. The authors sought to develop age-specific risk scores that predict high health care costs among an urban population. A retrospective cohort study was performed of children ages 1–18 years who received care at Fair Haven Community Health Care (FHCHC), a community health center in New Haven, Connecticut. Cost was estimated from charges in the electronic health record (EHR), which is shared with the only hospital system in the city. Using multivariable logistic regression models, independent predictors of being in the top decile of total charges during the 2017 calendar year were identified, drawing from covariates collected from the EHR prior to 2017. Random forest modeling was used to verify the feature importance of significant covariates and model performance from 2017 cost data were compared to those using 2018 cost data. Regression models were used to construct age-specific nomograms to predict cost. Among 8960 children who received care at FHCHC in the 18 months prior to 2017, covariate frequencies clustered in age groups 1–5 years, 6–11 years, and 12–18 years, so 3 age-specific models were constructed. Prior utilization variables predicted future costs, as did younger children who received specialty care and older children with behavioral health diagnoses. Final models for each age group had C statistics ≥0.68 using both 2017 and 2018 cost data. Prediction models can draw from elements accessible in the EHR to predict cost of ambulatory pediatric patients. Strategies to impact utilization among high-risk children are needed.