Using Re-inforcement Learning to Automatically Adapt a Remote Therapy Intervention (RTI) for Reducing Adolescent Violence Involvement

青少年暴力是一个关键的公共卫生问题。杀人罪是青少年死亡的主要原因(年龄：14- 20)，而且对非裔美国人人口的影响不成比例。城市电子政务是一个关键机遇 预防暴力，特别是每年有60万青少年因与暴力有关的伤害而寻求治疗。 在我们对城市急诊室暴力伤害青少年的纵向研究中，我们发现在两年内，37%的人返回 对于重复的暴力伤害，59%的人经历过枪支暴力，38%的人被捕，1%的人死亡。尽管 这一问题的重要性、减少急诊科就诊后再次发生暴力的策略尚未得到很好的研究。 鉴于我们之前的工作表明，基于理论的单次ED干预是有效的 减少低风险青少年中的暴力，这一疗法的应用扩大到解决更多的 跨多个会话的问题严重性，并通过包括护理管理得到增强，代表了潜在的 改变高风险暴力伤害青少年风险轨迹的有效方法。我们最近试行的 这一方法(S-RTI)得到了很好的接受，并解决了先前多期干预中发现的问题 (例如，运输)以及远程治疗递送(例如，电话)。虽然创新和前景看好， 这种S-放射治疗方法是资源密集型的，不能解决治疗反应的异质性。通过 相比之下，适应性治疗策略允许“及时”裁剪，在过多的 没有足够的干预措施，在减少使用昂贵资源的同时提高了成果。补强 学习是一个人工智能领域，它允许计算机系统从之前的成功中学习 这是一种很有前途的构建适应性“及时”干预措施的方法。在这项研究中，我们 建议测试我们的RTI的两个版本，一个由单个ED会话组成的标准RTI条件(S-RTI 之后是8次远程治疗，以及通过强化学习优化的自适应RTI版本(AI-RTI 在三个级别之间增加或降低治疗强度(即，远程治疗会话、自动两级 短信方式，仅限评估)基于患者对日常短信评估的响应。这个 具体目标是：1)完善和调整我们的RTI，以使用两个包(S-RTI；AI-RTI)交付；2)进行 一项三臂随机对照试验招募了900名寻求急诊护理的暴力伤害青少年(年龄：14-20岁)，以比较 S-RTI(n=300)、AI-RTI(n=400)和控制条件(n=200)；3)评估AI-RTI的适应性 RL算法通过比较第一个50%的参与者和第二个50%的过程变量(例如，参与， 乐于助人/讨人喜欢)主要结果(在4个月、8个月和12个月时评估)包括攻击、受害、 以及埃德因暴力伤害而累犯。次要结果包括药物使用、精神健康症状和 刑事司法介入。作为次要目标，我们将比较资源利用率(即成本/避免的事件) 对于积极干预的条件。鉴于社会弱势青年的暴力发生率上升， 尽管在获得服务方面存在差异，但拟议的研究有可能对公共卫生产生重大影响。

Youth violence is a key public health problem. Homicide is a leading cause of death among adolescents (age:14- 20) and disproportionately impacts African-American populations. Urban EDs are a critical opportunity for violence prevention, especially with >600,000 adolescents/year seeking treatment for violence-related injuries. In our longitudinal study of violently-injured adolescents in urban EDs, we found that within 2-years, 37% returned for a repeat violent injury, 59% experienced firearm violence, 38% were arrested, and 1% died. Despite the importance of the problem, strategies to decrease repeat violence after an ED visit have not been well studied. Given our prior work demonstrating that theoretically-based single session ED interventions are efficacious reducing violence among lower risk adolescents, the application of this therapy, expanded to address greater problem severity over multiple sessions and enhanced by including care management, represents a potentially efficacious approach for altering risk trajectories of higher-risk violently-injured adolescents. Our recent pilot of this approach (S-RTI) was well received and addressed problems identified in prior multisession interventions (e.g., transportation) with the addition of remote therapy delivery (e.g., phone). While innovative and promising, this S-RTI approach is resource intensive and does not address heterogeneity in treatment responses. By contrast, adaptive treatment strategies allow for “just-in-time” tailoring that provides a balance between too much and not enough intervention and enhances outcomes while reducing the use of costly resources. Reinforcement learning is an artificial intelligence domain that allows computer systems to “learn” from the success of prior treatments and is a promising approach to constructing adaptive “just-in-time” interventions. For this study, we propose to test two versions of our RTI, a standard RTI condition (S-RTI) comprised of a single ED session followed by 8 remote therapy sessions, and an adaptive RTI version (AI-RTI) optimized by reinforcement learning to step up or down the intensity of treatment between three levels (i.e., remote therapy sessions, automated two- way text messaging, assessment only) based on patient response to daily text message assessments. The specific aims are: 1) To refine and adapt our RTI for delivery using two packages (S-RTI; AI-RTI); 2) To conduct a 3-arm RCT enrolling 900 violently-injured adolescents seeking ED care (age:14-20) to compare the efficacy of S-RTI (n=300), AI-RTI (n=400), and a control condition (n=200); and, 3) To evaluate adaptability of the AI-RTI RL algorithm by comparing the first 50% of enrollees to the second 50% on process variables (e.g., engagement, helpfulness/likability). Primary outcomes (assessed at 4-, 8-, and 12-months) include aggression, victimization, and ED recidivism for violent injury. Secondary outcomes include substance use, mental health symptoms, and criminal justice involvement. As a secondary aim, we will compare resource utilization (i.e., costs/event averted) for the active intervention conditions. Given elevated rates of violence among socio-disadvantaged youth with disparities in access to services, the proposed study has the potential for significant public health impact.