Making Computerized Trauma Triage Decision Support Accurate and Trustworthy

摘要/摘要 创伤分诊经常发生在以时间和信息为特征的高压力环境中 对于做出相应决策来说不是最优的约束。这样的条件使得有必要 依靠对急救医务人员来说足够简单和直接的决策规则集 在提供急需的病人护理的同时，迅速执行。到目前为止，已发表的分流研究尚未 尽管努力优化创伤分诊流程，但仍实现了创伤分诊系统性能的目标。 目前的分诊系统可能无法实现这些目标。我们之前的工作表明，允许更多 复杂的规则和更详细的数据可以实现朝着这些目标迈出的重要一步。我们的长期目标是 建立智能、学习型的创伤计算机分诊决策支持(CTDS)系统，该系统由 信息丰富的环境，收集和处理院前数据，并有效地进行准确沟通 以及可以理解的分诊建议，以改善患者的预后。朝向这一目标的拟议步骤 将验证和扩展我们的初步结果，并评估人工智能生成的解释的复杂性 以提高这种CTDS系统的可信度。我们建议使用大规模的人口统计和 地理位置不同的数据集，以首先构建并定量评估多个复合体的性能 模特们。然后，我们建议评估这些复杂模型的群体公平性，并评估多重偏差 缓解策略，最后，我们建议与护理人员合作，设计通过算法生成的， 以EMS为导向的解释并评估这些解释的可信度。建议的项目是 创新，首先是因为它包含了似乎需要接近公布的准确性所需的复杂性 目标，同时评估解决与此相关的挑战的实用技术 复杂性。其次，它将通过处理机会和 新兴技术(例如，低成本、联网传感器)给院前带来的挑战 做决定。拟议的项目意义重大，因为减少误婚患者的数量可以 带来大量成本节约和死亡率降低，但目前的分诊系统可能无法实现 敏感度和特异度目标，甚至显著降低目前的失婚率。通过以下方式提高准确性 然而，复杂的模型可能不足以导致我们所设想的有影响力的变化。接受 如果已知偏见得到缓解，并且如果 推荐解释被认为是值得信赖的。

Abstract/Summary Trauma triage frequently occurs in high stress environments characterized by time and information constraints that are suboptimal for making consequential decisions. Such conditions have made it necessary to rely on decision-making rulesets that are simple and straightforward enough for emergency medical personnel to execute quickly while providing urgently needed patient care. To date, published triage studies have not achieved the goals for trauma triage system performance despite efforts to optimize the trauma triage process. Current triage systems may not be able to achieve these goals. Our prior work demonstrated that allowing more complex rules with more detailed data can achieve a significant step toward those goals. Our long-term aim is to build an intelligent, learning computerized trauma triage decision support (CTDS) system, that, aided by an information-rich environment, collects and processes prehospital data and effectively communicates accurate and understandable triage recommendations that improve patient outcomes. The proposed step toward this goal will validate and extend our preliminary results and assess the complexity of AI-generated explanations intended to improve the trustworthiness of such a CTDS system. We propose using a large demographically and geographically diverse data set to first build and quantitatively assess the performance of multiple complex models. We propose to then assess the group fairness of these complex models and evaluate multiple bias mitigation strategies, and lastly, we propose working with paramedics to both design algorithmically generated, EMS-oriented explanations and assess the trustworthiness of those explanations. The proposed project is innovative, first, because it embraces the complexity that appears to be required to approach published accuracy goals while simultaneously assessing practical techniques to address the challenges associated with that complexity. Second, it will help define a path forward for trauma triage by addressing opportunities and challenges that emerging technologies (e.g., low-cost, Internet-connected sensors) create for prehospital decision making. The proposed project is significant because reducing the number of mistriaged patients can result in substantial cost-savings and mortality reduction, but current triage systems may not be able to achieve sensitivity and specificity goals or even significantly reduce current mistriage rates. Improving accuracy through complex models, however, might not be enough to result in the impactful change we envision. The acceptance of such recommendations from such models is likely to improve if bias known to be mitigated and if recommendation explanations are seen as trustworthy.