Predicting food insecurity in a pediatric population using the electronic health record

More than 5 million children in the United States experience food insecurity (FI), yet little guidance exists regarding screening for FI. A prediction model of FI could be useful for healthcare systems and practices working to identify and address children with FI. Our objective was to predict FI using demographic, geographic, medical, and historic unmet health-related social needs data available within most electronic health records.

This was a retrospective longitudinal cohort study of children evaluated in an academic pediatric primary care clinic and screened at least once for FI between January 2017 and August 2021. American Community Survey Data provided additional insight into neighborhood-level information such as home ownership and poverty level. Household FI was screened using two validated questions. Various combinations of predictor variables and modeling approaches, including logistic regression, random forest, and gradient-boosted machine, were used to build and validate prediction models.

A total of 25,214 encounters from 8521 unique patients were included, with FI present in 3820 (15%) encounters. Logistic regression with a 12-month look-back using census block group neighborhood variables showed the best performance in the test set (C-statistic 0.70, positive predictive value 0.92), had superior C-statistics to both random forest (0.65, p < 0.01) and gradient boosted machine (0.68, p = 0.01), and showed the best calibration. Results were nearly unchanged when coding missing data as a category.

Although our models could predict FI, further work is needed to develop a more robust prediction model for pediatric FI.