Leveraging EHR data to evaluate key treatment decisions to prevent suicide-related behaviors

项目摘要/摘要 主要焦点：利用EHR数据(PAR-18-929)预防自杀相关行为(SRB)的研究 目的：使用增强型退伍军人健康管理局(VHA)的EHR数据来开发个性化 治疗规则(PTR)，帮助指导临床医生为精神病患者做出关键的治疗决定 目标是在未来12个月内减少SRB。 具体目标：我们将重点关注两个决定：初级保健医生对如何治疗患者的决定 向他们寻求常见精神障碍的帮助(CMD；“PCP研究”)；以及VHA的决定 自杀预防协调员关于是否将刚刚进行了非致命自杀尝试的患者送往医院或 把他们当做门诊病人对待(“SPC研究”)。两者都被认为是关键决策，没有全局最优 任何一种治疗路径都没有，对于如何在治疗方案中做出决定几乎没有指导。 研究设计：我们将采用前瞻性观察设计。PCP研究将以EHR数据为基础 在2010-2016年间，大约有583,000名VHA患者前往PCP寻求CMD帮助。事故探访 将被定义为患者在过去12个月内没有接受过其他CMD治疗。五人组 广泛的PCP治疗选择包括药物治疗、转诊至心理治疗、药物治疗加 心理治疗，药物治疗加上基于测量的协作护理，以及转介给精神病学家。 结果要么是SRB(主要结果，要么是自杀死亡，要么是行政记录 非致命性自杀未遂)在接下来的12个月内或在同一时间内有自杀倾向的精神病住院治疗 随访期(次要结果)。这些结果发生在2010-2016年间12,292次事件查访之后。 SPC的研究将基于2010-2016年间完成的67,196份VHA自杀行为报告 非致命性VHA门诊自杀未遂。这些病例中大约有一半入院治疗，其他人接受治疗。 作为门诊病人。在接下来的12个月里，19,829例患者发生了重复的SRB。 方法：使用平衡基线协变量的最佳实践方法对非随机变量进行调整 治疗方案中的任务分配。基准协变量将包括：以前的电子健康记录数据；电子健康记录数据 局部治疗决策，包括用自然语言从临床记录中提取的信息 处理；患者地址的小区域地理编码数据；来自LexisNexis Social的个人级别数据 关于患者财务、就业、婚姻状况和刑事司法的健康决定因素数据库 参与；以及关于治疗临床医生的先前实践模式和实践的信息-资源 治疗设置。将使用一种前沿的集成机器学习方法来分析这些权重 用于制定PTRS的数据。2010-2016年数据交叉验证和2017-2018年数据验证(不可用 (直到研究的第三年)将被用来估计临时技术报告的样本外绩效。

PROJECT SUMMARY/ABSTRACT Key focus: Research leveraging EHR data (PAR-18-929) to prevent suicide-related behaviors (SRBs) Objectives: To use augmented Veterans Health Administration (VHA) EHR data to develop Personalized Treatment Rules (PTRs) to help guide clinicians in making key treatment decisions for mentally ill patients aimed at reducing SRBs over the next 12 months. Specific aims: We will focus on two decisions: the decision of primary care physicians on how to treat patients coming to them for help with common mental disorders (CMD; “the PCP study”); and the decision of VHA Suicide Prevention Coordinators on whether to hospitalize patients who just made nonfatal suicide attempts or treat them as outpatients (“the SPC study”). Both are recognized as critical decisions, with no globally optimal treatment path for either and little guidance on how to decide among the treatment options. Research design: We will use a prospective observational design. The PCP study will be based on EHR data for the roughly 583,000 incident PCP visits of VHA patients for help with a CMD in 2010-2016. An incident visit will be defined as where the patient had not received other CMD treatment in the prior 12 months. The five broad PCP treatment options are pharmacotherapy, referral to psychotherapy, pharmacotherapy plus psychotherapy, pharmacotherapy plus measurement based collaborative care, and referral to a psychiatrist. The outcomes will be either an SRB (the primary outcome, either suicide death or administratively-recorded nonfatal suicide attempt) over the next 12 months or psychiatric hospitalization with suicidality over the same follow-up period (the secondary outcome). These outcomes occurred after 12,292 2010-2016 incident visits. The SPC study will be based on the 67,196 2010-2016 VHA Suicide Behavior Reports completed after a nonfatal VHA outpatient suicide attempt. Roughly half of these cases were hospitalized and the others treated as outpatients. A repeat SRB occurred over the next 12 months for 19,829 of these cases. Methods: A best-practice method of balancing baseline covariates will be used to adjust for nonrandom assignment across treatment options. Baseline covariates will include: prior EHR data; EHR data available for the focal treatment decision, including information abstracted from clinical notes with natural language processing; small-area geocode data for patient addresses; individual-level data from the LexisNexis Social Determinants of Health Database on patient finances, employment, marital status, and criminal justice involvement; and information about prior practice patterns of treating clinicians and practices-resources of treatment settings. A cutting-edge ensemble machine learning method will be used to analyze these weighted data to develop PTRs. Cross-validation in the 2010-2016 data and validation in 2017-2018 data (not available until the third year of the study) will be used to estimate out-of-sample performance of the PTRs.