Multi-site External Validation and Improvement of a Clinical Screening Tool for Future Firearm Violence

摘要 在临床环境中的干预，如急诊科(ED)，是一个机会 预防人际枪支暴力，特别是在青少年中，他们的人际枪支暴力 不成比例地影响。成功的临床干预的关键前提是准确地评估风险， 确保对稀缺资源的合理分配；提供缺少的先决条件是 建议的工作。与传统的推论统计模型相比，机器学习方法是 其特点是强调前瞻性预测，并在几个领域加强了临床预测， 包括心脏病、癌症诊断和预后、创伤后应激障碍、自杀风险和药物使用等。 然而，除了安全评分--由目前的调查小组制定--机器学习 还没有利用各种方法来前瞻性地预测枪支暴力。在这项拟议的工作中，我们的研究 目标有两个：1)通过确定其预测武器的能力来外部验证安全分数 在未来一年内在新的数据集上涉及暴力；以及2)通过进行 弹性净惩罚Logistic回归四种机器学习方法的比较分析 随机森林、支持向量机和增强(集成)方法。通过这种方式，我们正在回应 目标一：研究有助于促进创新和有前途的机会的发展，以提高 安全和防止与枪支有关的伤害、死亡和犯罪。这种方法是创新的，因为它建立在 将机器学习方法应用于枪支暴力预测是唯一的工作，这是一个很有希望的机会 防止火器伤害，因为这将a)明确衡量未来的枪支暴力风险；以及b) 与以前的任何研究不同，根据风险因素的前瞻性预测能力来表征风险因素的影响。因此，这就是 研究既将确定最需要干预的个人，也将指出潜在的可修改 预测因素。以一种概括性的方式适当地解决这个研究问题需要当代的 1)重点关注高需求但范围广泛的研究人群的数据集；2)全面的基准衡量标准， 为预测提供广泛的基础；以及3)地理变异性(中西部、西海岸和东海岸) 增强了泛化能力。因此，我们将从三个截然不同的城市急救队招募1500名年龄在18-24岁的年轻人 地点-弗林特、费城和西雅图-并管理涵盖以下几个领域的基线调查 未来暴力的潜在危险因素，并在6个月和12个月时与这些青少年进行跟踪，以确定 主要结果--(作为受害者或施暴者)卷入枪支暴力--以及次要结果： 高危持枪行为、非持枪暴力、暴力伤害。因为这项工作需要一个前瞻性的 纵向研究，我们正在申请方案B。这项工作将为未来涉及 制定和测试针对人际枪支暴力的干预措施，包括确定潜在的高风险 利用可修改的预测因素，并确定最需要干预的青年。

ABSTRACT Interventions in clinical settings, such as the emergency department (ED), are an opportunity for interpersonal firearm violence prevention, particularly among youth, whom interpersonal firearm violence disproportionately affects. A crucial prerequisite to successful clinical interventions is an accurate gauge of risk, to ensure the judicious allocation of scarce resources; providing that missing prerequisite is the primary goal of the proposed work. Machine learning methods, in contrast to traditional inferential statistical models, are distinguished by their emphasis on prospective prediction, and have enhanced clinical prediction in several fields, including heart disease, cancer diagnosis and outcomes, PTSD, suicide risk, and substance use, among others. Yet, with the exception of the SAFETY score—developed by the current investigative team—machine learning methods have not been leveraged to prospectively predict firearm violence. In this proposed work our research objectives are two-fold: 1) Externally validate the SAFETY score by determining its ability to predict firearm violence involvement within the next year on a new data set; and 2) Improve the SAFETY score by conducting a comparative analysis of four powerful machine learning methods: elastic net penalized logistic regression, random forests, support vector machines, and boosting (ensemble) methods. In this way, we are responding to Objective One: Research to help inform the development of innovative and promising opportunities to enhance safety and prevent firearm-related injuries, deaths, and crime. This approach is innovative because it builds upon the only work to apply machine learning methods to firearm violence prediction, and it is a promising opportunity to prevent firearm injuries because it will a) provide an explicit gauge of future firearm violence risk; and b) characterize risk factor effects in terms of their prospective prediction ability, unlike any prior research. Thus this research will both identify individuals in most need of intervention, and also point to potentially modifiable predictive factors. Properly addressing this research question in a generalizable way requires a contemporary data set with 1) a focus on a high-need, yet broad, study population; 2) comprehensive baseline measures that provide a broad basis for prediction; and 3) geographic variability (Midwest, West Coast, and East Coast) that enhances generalizability. Thus, we will recruit 1,500 youth age 18-24 from urban EDs in three broadly different locales—Flint, Philadelphia, and Seattle—and administer a baseline survey covering several domains of potential risk factors for future violence, and follow up with those youth at 6- and 12-months to ascertain the primary outcome—firearm violence involvement (as victim or perpetrator)—as well as the secondary outcomes: high-risk firearm behaviors, non-firearm violence, and violent injury. Because this work requires a prospective longitudinal study, we are applying for Option B. This work will lay the ground for future research involving the development and testing of interventions for interpersonal firearm violence both by identifying potential high- leverage modifiable predictive factors, and by identifying youth most in need of intervention.