(893) A Machine Learning Approach to Identify Child Opportunity Predictors Associated With Firearm Injury

Introduction: Firearm injury is the leading cause of death in children and adolescents in the United States. Exposure to gun violence is influenced by social determinants of health. We previously showed Child Opportunity Index (COI) was associated with pediatric firearm injury in Philadelphia, Pennsylvania. Using a machine learning approach, we aimed to describe the relative importance of COI subdomains in predicting pediatric firearm injury, hypothesizing that there are modifiable structural factors associated with areas with high burden of injury.

Methods: Using publicly available shooting data, we identified shooting victims < 21 years of age from 2015 to June 2024 in Philadelphia. The population adjusted incidence by census tract was calculated using 2017-2021 ACS 5-year population estimates. A random forest regressor model to predict census-level shooting incidence was generated using 14 nationally normed COI 3.0 subdomain scores by census tract. We used a training split of 70:30 with 5-fold cross validation. Minimum RMSE values were used to select the best performing model, which was then applied to the test dataset; R2 was calculated and a scatter plot of actual versus predicted with a fitted regression line was generated to assess fit and calibration. Relative variable (COI subdomain) importance –taken to be the mean increase in node purity—was then determined. Analysis was performed using R (version 4.4.0; R Core Team, 2024) and RStudio (version 2022.2.2.485; RStudio Team, 2024).

Results: There were 3,462 pediatric firearm injuries during the study period, with 557 fatalities (mortality rate 16%). The best performing model had R2 =0.72. COI subdomains were ranked by variable importance with 1) educational resources 2) housing resources 3) economic resources and 4) concentrated socioeconomic inequity as highest predictors of pediatric firearm injury.

Conclusions: A machine learning approach to predicting pediatric firearm injury using 14 COI subdomains yielded a model with R2 =0.72. COI subdomains of 1) educational resources 2) housing resources 3) economic resources and 4) concentrated socioeconomic inequity were most predictive of pediatric firearm injury. Further investigation of these subdomains is need to identify potential modifiable structural factors associated with firearm injury.