Improving safety and quality of emergency care using machine learning-based clinical decision support at triage

摘要 目前，美国各地的急诊科（艾德）分诊系统导致多达三分之一的患者出现错误分诊。 患者接触，恶化艾德拥挤，并导致护理延迟和差异。早期研究 显示，通过优先排序，分流预测模型的性能优于大多数急诊科使用的主观分流系统， 病人，尽管仍有很大差距。早期的模型没有：1）预测医院以外的结果 入院，即使80-90%的艾德访视不会导致入院; 2）包括儿科患者， 即使儿科患者占就诊人数的25%; 3）在设计或评价中考虑健康公平性 预测模型;或4）研究对关键患者安全和质量指标的影响，例如 在乎我们的提案解决了这些未满足的需求，并回应了两个AHRQ特别强调通知，HS- 21-014（促进健康公平的健康服务研究）和HS-22-004（数字医疗保健研究） 安全）。我们的研究团队已经完成了重要的初步分析，包括：1）研究队列建立超过 在我们的卫生系统中，21个ED中有600万艾德患者; 2）评估 研究环境中的分诊;以及3）开发机器学习模型，以预测患者的紧急程度和资源 需要分诊。在目标1中，我们将改进分类模型，预测：1）危重病; 2）住院; 3)快速通道资格（需要<2个资源，无住院或严重结局）。我们将测量 算法偏差和探索策略，以提高分流模型预测的公平性。在目标2中，我们将绘制 将每个结果的概率阈值划分为临床相关的分诊类别建议。我们将使用一个 人为因素框架和重要的利益相关者参与设计，构建和评估临床医生- 临床决策支持（CDS）。最后，我们会将CDS纳入电子健康记录， 作为标准分诊工作流程的一部分，有效显示每个结果的个性化风险预测。在目标3中， 我们将在一项务实的、逐步楔入的随机分组试验中，在真实的时间内评估CDS的影响。 21个医院急诊室和一个独立的急诊室。我们的主要结果是：1） 危重患者; 2）适当早期识别快速通道合格患者; 3）艾德住院时间。在 此外，为了测试我们的模型中的股权驱动校准，我们将评估种族，性别和 社会经济地位的主要结果。我们的次要结果将是CDS的覆盖率、采用率和 实施.在成功完成拟议的研究后，我们预计将证明， 一种使用先进预测分析的新型床旁数字技术可以带来更安全、更高的 质量和更公平的护理。

Abstract Current emergency department (ED) triage systems across the U.S. lead to mis-triage in up to one third of patient encounters, worsening ED crowding and contributing to delays and disparities in care. Early studies showed triage prediction models outperformed the subjective triage systems used in most EDs by prioritizing sicker patients, although significant gaps remain. Early models did not: 1) predict outcomes other than hospital admission, even though 80-90% of ED visits do not result in hospital admission; 2) include pediatric patients, even though pediatric patients contribute up to 25% of visits; 3) consider health equity in design or evaluation of prediction models; or 4) study impacts on key patient safety and quality measures, such as timeliness of care. Our proposal addresses these unmet needs and responds to two AHRQ Special Emphasis Notices, HS- 21-014 (Health Services Research to Advance Health Equity) and HS-22-004 (Research on Digital Healthcare Safety). Our study team has completed significant preliminary analyses, including: 1) study cohort build of over 6 million ED encounters across the 21 EDs in our health system; 2) assessment of significant limitations of triage in study setting; and 3) development of machine-learning models to predict patient acuity and resource needs at triage. In Aim 1, we will refine triage models that predict: 1) critical illness; 2) hospital admission; and 3) fast-track eligibility (<2 resources needed, no hospital admission or critical outcomes). We will measure algorithm biases and explore strategies to improve equity in triage model predictions. In Aim 2, we will map probability thresholds for each outcome into clinically relevant triage category recommendations. We will use a human factors framework and significant stakeholder engagement to design, build, and evaluate clinician- facing triage clinical decision support (CDS). Lastly, we will build the CDS into our electronic health record to efficiently display personalized risk predictions for each outcome as part of standard triage workflows. In Aim 3, we will assess the impact of the CDS in real time in a pragmatic, step-wedged cluster randomized trial across 21 hospital-based EDs and one free-standing ED. Our primary outcomes will be: 1) timeliness of care for critically ill patients; 2) appropriate early identification of fast-track eligible patients; and 3) ED length of stay. In addition, to test the equity-driven calibrations in our models, we will assess for bias by race, gender, and socioeconomic status among primary outcomes. Our secondary outcomes will be CDS reach, adoption, and implementation. Upon successful completion of the proposed research, we expect to demonstrate the extent to which a novel point-of-care digital technology that uses advanced predictive analytics can lead to safer, higher quality, and more equitable care.