A machine-learning prediction model to identify risk of firearm injury using electronic health records data.

Abstract

Importance: Firearm injuries constitute a public health crisis. At the healthcare encounter level, they are, however, rare events.

Objective: To develop a predictive model to identify healthcare encounters of adult patients at increased risk of firearm injury to target screening and prevention efforts.

Materials and methods: Electronic health records data from Kaiser Permanente Southern California (KPSC) were used to identify healthcare encounters of patients with fatal and non-fatal firearm injuries, as well as healthcare visits of a sample of matched controls during 2010-2018. More than 170 predictors, including diagnoses, healthcare utilization, and neighborhood characteristics were identified. Extreme gradient boosting (XGBoost) and a split sample design were used to train and test a model that predicted risk of firearm injury within the next 3 years at the encounter level.

Results: A total of 3879 firearm injuries were identified among 5 288 529 KPSC adult members. Prevalence at the healthcare encounter level was 0.01%. The 15 most important predictors included demographics, healthcare utilization, and neighborhood-level socio-economic factors. The sensitivity and specificity of the final model were 0.83 and 0.56, respectively. A very high-risk group (top 1% of predicted risk) yielded a positive predictive value of 0.14% and sensitivity of 13%. This high-risk group potentially reduces screening burden by a factor of 11.7, compared to universal screening. Results for alternative probability cutoffs are presented.

Discussion: Our model can support more targeted screening in healthcare settings, resulting in improved efficiency of firearm injury risk assessment and prevention efforts.