Development and Spatial Validation of a Random Forest Prediction Model for Firearm-Related Injury Risk in Chicago Census Tracts

Abstract Firearm-related injury is a public health crisis, particularly in cities like Chicago, where incidents are spatially concentrated in neighborhoods characterized by racial segregation, socioeconomic disadvantage, and environmental neglect. While traditional random forest models can identify important neighborhood-level predictors within these broad categories, they ignore spatial patterning, where nearby areas exhibit similar levels of risk, thereby overlooking potential clustering or diffusion of firearm injury across space. As a result, non-spatial random forest models (NSRFM) are prone to biased estimates and may have limited generalizability across different locations. To address this gap, we applied a spatial random forest model (SRFM) to analyze non-fatal firearm-related injuries across Chicago census tracts from 2020 to 2024, using a comprehensive set of social, natural, and built environmental predictors. Results showed that accounting for spatial patterns significantly improved predictive performance: the spatial random forest achieved a pseudo-R² of 0.98 and eliminated residual spatial autocorrelation. Spatial cross-validation revealed considerable variation in model transferability across neighborhoods (median normalized RMSE = 0.76; range: 0.52–3.03). The most important variables included racial composition, childhood lead exposure, social vulnerability, traffic volume, tree canopy cover, and disinvestment in single-family homes. Our results demonstrate the value of spatially explicit models for uncovering hidden neighborhood-level risk and guiding place-based violence prevention efforts that account for the most important socio-environmental determinants.