A novel machine-learning tool to identify community risk for firearm violence: The Firearm Violence Vulnerability Index

Abstract

Background: Firearm violence in the United States is a public health crisis, but accessing accurate firearm assault data to inform prevention strategies is a challenge. Vulnerability indices have been used in other fields to better characterize and identify at-risk populations during crises, but no tool currently exists to predict where rates of firearm violence are highest. We sought to develop and validate a novel machine-learning algorithm, the Firearm Violence Vulnerability Index (FVVI), to forecast community risk for shooting incidents, fill data gaps, and enhance prevention efforts.

Methods: Open-access 2015 to 2022 fatal and nonfatal shooting incident data from Baltimore, Boston, Chicago, Cincinnati, Los Angeles, New York City, Philadelphia, and Rochester were merged on census tract with 30 population characteristics derived from the 2020 American Community Survey. The data set was split into training (80%) and validation (20%) sets; Chicago data were withheld for an unseen test set. XGBoost, a decision tree-based machine-learning algorithm, was used to construct the FVVI model, which predicts shooting incident rates within urban census tracts.

Results: A total of 64,909 shooting incidents in 3,962 census tracts were used to build the model; 14,898 shooting incidents in 766 census tracts were in the test set. Historical third grade math scores and having a parent jailed during childhood were population characteristics exhibiting the greatest impact on FVVI's decision making. The model had strong predictive power in the test set, with a goodness of fit ( D2 ) of 0.77.

Conclusion: The Firearm Violence Vulnerability Index accurately predicts firearm violence in urban communities at a granular geographic level based solely on population characteristics. The Firearm Violence Vulnerability Index can fill gaps in currently available firearm violence data while helping to geographically target and identify social or environmental areas of focus for prevention programs. Dissemination of this standardized risk tool could also enhance firearm violence research and resource allocation.

Level of evidence: Prognostic and Epidemiological; Level IV.