Optimizing Clinical Decision Support Alerts Using Explainable Artificial Intelligence (XAI)

项目摘要 在过去的十年里，联邦政府已经花费了超过340亿美元用于有意义地使用电子邮件， 电子健康记录（EHR）。然而，临床决策支持（CDS）警报的接受率， EHR的组成部分，小于10%。大量的低相关性警报（例如， 心脏复苏）不仅增加了临床医生的负担，而且可能导致警觉疲劳的开始， 从而导致忽略重要警报并对患者安全构成严重威胁。目前，警报 主要通过人工审核和收集用户反馈进行改进。然而，这些方法是劳动力- 密集的，不允许从全面的方面及时分析用户对警报的响应。的 警报日志数据量很大，范德比尔特大学医学中心在 2020.迫切需要利用警报日志和电子健康记录中的数据来开发数据驱动的流程， 生成改进警报逻辑或改进临床流程的建议。 为了解决这一差距，我建议使用可解释人工智能（XAI）结合偏见缓解技术， niques构建预测模型，全面了解用户对警报的响应， 生成负责任的建议，以改善警报的原始逻辑。该项目分为两个阶段 有三个具体目标。在K99阶段，我将接受多学科专家团队的指导，学习 最新的XAI和偏差缓解技术，以及CDS评估和管理，并实现 以下两个目标：目标1）开发一个基于标准的特征分类法，这些特征影响用户对CDS的响应 警报和目标2）开发数据驱动的流程，以生成使用XAI改进警报标准的建议 接近。然后，我将过渡到独立研究阶段R 00，以实现目标3）评估生成 建议采用混合方法设计。通过这项研究，我希望提供一个基于标准的 影响用户对警报响应的功能分类，这是一个创新的数据驱动流程，能够生成 改善警报的建议。我还将提出一套经过专家验证的建议。这项研究可能意味着- 有助于CDS管理和临床流程的改进。 我的职业发展计划和拟议的研究与我目前的技能和经验相一致， CDS和机器学习基于我的导师团队的互补专业知识，我将培养能力- 在四个领域：CDS，信息学方法，实施科学，职业发展和专业- 把它交给一个独立的研究者。总的来说，这个项目可以帮助我开展一项独立的研究， 开发可解释的智能CDS工具，以提高患者安全性，提供标准化护理， 建立公平、高效的医疗体系。

PROJECT SUMMARY Over the past decade, the federal government has spent more than $34 billion on the meaningful use of elec- tronic health records (EHRs). However, the acceptance rate for clinical decision support (CDS) alerts, a critical component of EHRs, is less than 10%. The large number of low relevance alerts (e.g. a weight loss alert during a cardiac resuscitation) not only increases the burden on clinicians, but can lead to the onset of alert fatigue, resulting in the neglect of important alerts and posing a serious threat to patient safety. Currently, alerts are improved primarily through manual review and by collecting user feedback. However, these methods are labor- intensive and do not allow for a timely analysis of user responses to alerts from a comprehensive aspect. The amount of alert log data is large, Vanderbilt University Medical Center generated over 3 million alert firings in 2020. There is an urgent need to utilize the data from the alert log and EHR to develop a data-driven process to generate suggestions for refining alert logic or improving clinical processes. To address this gap, I propose to use explainable artificial intelligence (XAI) combined with bias mitigation tech- niques to build predictive models that comprehensively learn user responses to alerts and in turn automatically generate responsible suggestions to improve the original logic of alerts. This project is divided into two phases with three specific aims. In the K99 phase, I will be mentored by a multidisciplinary team of experts to learn the latest XAI and bias mitigation techniques, as well as CDS evaluation and management, and to achieve the following two aims: Aim 1) Develop a standard-based taxonomy of features that affect user response to CDS alerts and Aim 2) Develop a data-driven process to generate suggestions for improving alert criteria using XAI approaches. I will then transition to the independent research phase R00 to achieve Aim 3) Evaluate generated suggestions using a mixed-methods design. Throughout this research, I expect to provide a standards-based taxonomy of features that affect user response to alerts, an innovative data-driven process capable of generating suggestions to improve alerts. I will also produce a set of expert-validated suggestions. This study could signif- icantly contribute to the CDS management and clinical processes improvements. My career development plan and the proposed research are aligned with my current skills and experiences in CDS and machine learning. Based on complementary expertise from my mentor team, I will develop competen- cies in four areas: CDS, informatics methods, implementation science, and career development and profession- alism to transfer to an independent researcher. Overall, this project can help me launch an independent research career in developing explainable, intelligent CDS tools to improve patient safety, provide standardized care, and promote an equitable and efficient healthcare system.