Newgard Craig D, Babcock Sean, Malveau Susan, Lin Amber, Goldstick Jason, Carter Patrick, Cook Jennifer N B, Song Xubo, Wei Ran, Salvi Apoorva, Fallat Mary E, Kuppermann Nathan, Jenkins Peter C, Fein Joel A, Mann N Clay
Center for Policy and Research in Emergency Medicine, Department of Emergency Medicine, Oregon Health & Science University, Portland, OR.
Institute for Firearm Injury Prevention, Injury Prevention Center, Department of Emergency Medicine, University of Michigan School of Medicine, Ann Arbor, MI.
Pediatr Emerg Care. 2025 Mar 1;41(3):195-202. doi: 10.1097/PEC.0000000000003314. Epub 2024 Dec 12.
Among children transported by ambulance across the United States, we used machine learning models to develop a risk prediction tool for firearm injury using basic demographic information and home ZIP code matched to publicly available data sources.
We included children and adolescents 0-17 years transported by ambulance to acute care hospitals in 47 states from January 1, 2014 through December 31, 2022. We used 96 predictors, including basic demographic information and neighborhood measures matched to home ZIP code from 5 data sources: EMS records, American Community Survey, Child Opportunity Index, County Health Rankings, and Social Vulnerability Index. We separated children into 0-10 years (preadolescent) and 11-17 years (adolescent) cohorts and used machine learning to develop high-specificity risk prediction models for each age group to minimize false positives.
There were 6,191,909 children transported by ambulance, including 21,625 (0.35%) with firearm injuries. Among children 0-10 years (n = 3,149,430 children, 2,840 [0.09%] with firearm injuries), the model had 95.1% specificity, 22.4% sensitivity, area under the curve 0.761, and positive predictive value 0.41% for identifying children with firearm injuries. Among adolescents 11-17 years (n = 3,042,479 children, 18,785 [0.62%] with firearm injuries), the model had 94.8% specificity, 39.0% sensitivity, area under the curve 0.818, and positive predictive value 4.47% for identifying patients with firearm injury. There were 7 high-yield predictors among children and 3 predictors among adolescents, with little overlap.
Among pediatric patients transported by ambulance, basic demographic information and neighborhood measures can identify children and adolescents at elevated risk of firearm injuries, which may guide focused injury prevention resources and interventions.
在美国通过救护车转运的儿童中，我们使用机器学习模型，利用基本人口统计信息和与公开可用数据源匹配的家庭邮政编码，开发一种枪支伤害风险预测工具。
我们纳入了2014年1月1日至2022年12月31日期间在47个州通过救护车转运至急诊医院的0至17岁儿童和青少年。我们使用了96个预测因素，包括基本人口统计信息以及从5个数据源（紧急医疗服务记录、美国社区调查、儿童机会指数、县卫生排名和社会脆弱性指数）与家庭邮政编码匹配的邻里指标。我们将儿童分为0至10岁（青春期前）和11至17岁（青少年）队列，并使用机器学习为每个年龄组开发高特异性风险预测模型，以尽量减少假阳性。
共有6,191,909名儿童通过救护车转运，其中21,625名（0.35%）有枪支伤害。在0至10岁的儿童中（n = 3,149,430名儿童，2,840名[0.09%]有枪支伤害），该模型在识别有枪支伤害的儿童时，特异性为95.1%，敏感性为22.4%，曲线下面积为0.761，阳性预测值为0.41%。在11至17岁的青少年中（n = 3,042,479名儿童，18,785名[0.62%]有枪支伤害），该模型在识别有枪支伤害的患者时，特异性为94.8%，敏感性为39.0%，曲线下面积为0.818，阳性预测值为4.47%。儿童中有7个高收益预测因素，青少年中有3个预测因素，重叠较少。
在通过救护车转运的儿科患者中，基本人口统计信息和邻里指标可以识别出枪支伤害风险较高的儿童和青少年，这可能有助于指导有针对性的伤害预防资源和干预措施。
