A Machine Learning Health System to Integrate Care for Substance Misuse and HIV Treatment and Prevention among Hospitalized Patients

摘要 药物滥用（SM）使人们面临感染艾滋病毒的风险。检测和处理风险的护理系统 艾滋病毒的感染或传播是孤立的。急诊室和医院提供了 在SM患者中筛查、检测和治疗HIV风险的独特机会。邻近 在芝加哥，拉什大学医学中心， 2017年成立了物质使用干预小组（SUIT）。SUIT服务尝试筛选所有 对SM住院患者进行干预，并根据低、中、高风险采取减少伤害的模式; 然而，忙碌的环境和病人疾病的尖锐性和严重性限制了普遍筛查率， 在决定筛查哪些患者时，容易产生隐性偏见。自动化临床决策 使用监督机器学习（ML）训练的支持工具可以减轻这些筛选负担。机器 学习型卫生系统方法利用EHR数据，包括临床、社会和行为决定因素 在结构化数据字段和临床笔记中捕获-非结构化数据通常无法用于预测 分析学ML HIV风险分类器可以识别具有SM和HIV风险的患者，并提醒提供者评估 适当的药物和护理，以预防或治疗艾滋病毒。到目前为止，还没有筛选工具， 开发和验证，以评估SM患者的艾滋病毒风险。这个飞行员的目标是开发，训练， 并测试可互操作的ML分类器，以确定患有以下疾病的患者中HIV传播或感染的风险： SM并评估其实时性能。目标1是开发、训练和测试一个ML分类器， 确定HIV感染或传播风险的敏感性（≥ 0.8）和特异性（≥0.8）， 滥用药物在2017-2019年SM患者的患者级数据的源队列中 （N= 23，817），我们将使用基于规则的方法和疾病控制中心的HIV风险指南来识别 这些患者遇到的诊断病例，如衣原体，与艾滋病毒传播有关（6%， n= 1，300）。利用倾向评分匹配，我们将匹配非病例（1：2）并进行手动图表 注释，以验证或重新分类病例和非病例，并建立参考数据集 （n= 3 900）。通过标记的案例和非案例，我们将对参考数据集进行分区，以训练和测试三个 监督学习ML模型我们将根据标准指标选择性能最佳的模型，如 C-统计量。目标2是将目标1中的最佳性能模型集成到Rush EHR基础设施中进行测试 预测效度在真实的时间，前瞻性。我们预计ML分类器可以识别50%以上的HIV风险 案例（9-10%）比我们的基于规则的方法，我们将研究分类器的效果，并衡量数量 在中断的时间序列中，在12个一个月的时间点上识别出的风险病例数。这个ML分类器是第一个 迈向一个适当的、可扩展的、可互操作的学习卫生系统干预措施， 将预防和治疗纳入SM住院患者的护理。

ABSTRACT Substance misuse (SM) puts persons at risk for HIV. Systems of care for the detection and treatment of risks for HIV acquisition or transmission among SM populations are siloed. The acute-care, hospital setting offers unique opportunities for screening, testing, and treatment of HIV risks among patients with SM. Adjacent to communities with the highest number of heroin overdoses in Chicago, Rush University Medical Center launched the Substance Use Intervention Team (SUIT) in 2017. The SUIT service attempts to screen all hospitalized patients for SM and intervenes with a harm reduction model based on low, medium, and high risk; however, the busy setting and acuity and severity of patients’ illnesses limit universal screening rates and facilitate implicit biases in making determinations about which patients to screen. Automated, clinical decision support tools trained with supervised machine learning (ML) can relieve these screening burdens. A machine learning health system approach leverages EHR data, including clinical, social, and behavioral determinants captured in structured data fields and in clinical notes – unstructured data typically unavailable for predictive analytics. A ML HIV risk classifier can identify patients with SM and HIV risk and alert providers to evaluate appropriateness for medication and care to prevent or treat HIV. To date, no screening tool has been developed and validated to assess for HIV risk among persons with SM. This pilot’s goal is to develop, train, and test an interoperable ML classifier to identify risk for HIV transmission or acquisition among patients with SM and assess its real-time performance. Aim 1 is to develop, train, and test a ML classifier with high sensitivity (≥0.8) and specificity (≥0.8) to identify risk for HIV acquisition or transmission among patients with substance misuse. Within the source cohort of encounter-level data of patients with SM between 2017-2019 (N=23,817), we will use a rule-based method and Centers for Disease Control HIV risk guidelines to identify as cases those patient encounters with diagnoses, such as Chlamydia, associated with HIV transmission (6%, n=1,300). Utilizing propensity score matching we will match non-cases (1:2) and conduct manual chart annotation in order to verify or re-classify cases and non-cases and to establish the reference dataset (n=3,900). With labeled cases and non-cases, we will partition the reference dataset to train and test three supervised learning ML models. We will select the best performing model based on standard metrics, like the C-statistic. Aim 2 is to integrate the best performing model from Aim 1 into the Rush EHR infrastructure to test predictive validity in real time, prospectively. As we expect the ML classifier to identify 50% more HIV risk cases (9-10%) than our rule-based method, we will study the effects of the classifier and measure the number of risk cases identified over 12 one-month time points in an interrupted time series. This ML classifier is the first step toward an appropriate, scalable, and interoperable learning health system intervention that integrates HIV prevention and treatments into care for hospitalized patients with SM.