Aligning Patient Acuity with Intensity of Care after Surgery

摘要 这项提议的一个关键目标是为候选人提供必要的培训和发展所需的资源 使用电子健康记录进行大规模、多机构信息学研究的专业知识和经验 数据和机器学习，以开发临床决策支持工具。这项建议是针对候选人的 长期的职业目标是成为一名在设计和实施方面具有专业知识的独立外科医生-科学家 机器学习系统，以增强临床决策。为了实现这一目标，候选人和 导师建议对术后“患者敏锐度”(即危重疾病和死亡的风险)进行系统调查。 和“护理强度”(即，生命体征和实验室测量的分诊目的地和频率)。之后 大手术、错位的患者敏锐度和护理强度可能会导致可预防的伤害和不适当的 资源使用，仅在美国每年就影响大约1500万例住院手术。当高敏锐度 患者接受低强度护理，术后并发症可进展为危重疾病和心脏骤停。 为低视力患者提供高强度护理的价值很低，并可能通过不必要的方式造成伤害 治疗。很难系统地解决这些问题，因为没有经过验证的、统一的 “护理强度”的定义。本应用程序的总体目标是了解护理强度决策 并将其与临床表型和结果进行匹配，利用这一知识来 生成精确、自主的决策支持工具。这个应用程序的中心假设是 不适当的术后护理强度是常见的，可预见的，并与增加的短期和 长期死亡率、发病率和医院费用。这项工作的基本原理是将电子健康 记录数据、机器学习和临床领域的专业知识提供了了解术后情况的机会 护理强度决策，并开发能够优化临床结果的决策支持工具 资源使用。本提案的具体目的是：(1)制定和验证术后护理强度 定义，(2)开发和验证可解释的、可操作的敏感度评估，以阐明决策空间， (3)识别和预测术后护理强度表型。这项拟议的研究具有重要意义 因为它解决了一个每年影响数百万患者的问题，并与潜在的 可预防的伤害和次优的资源利用。这种方法是创新的，因为候选人和导师 不知道任何先前试图对术后护理强度进行分类和判定的尝试，并理解 接受不充分或过度护理的患者的表型和特征。在颁奖期内， 候选人将申请NIH-R01研究员发起的奖项，以表彰其预期的临床实施 包含经过验证的护理强度定义和知识的可解释、可操作的决策支持工具 从表型聚类获得，最初处于静默数据收集期，之后是活跃期 向临床医生提供模型输出和临床可操作的建议。

ABSTRACT A key aim of this proposal is to equip the candidate with the training and resources necessary to develop expertise and experience in large-scale, multi-institutional informatics research using electronic health record data and machine learning to develop clinical decision-support tools. This proposal builds toward the candidate’s long-term career goal of becoming an independent surgeon-scientist with expertise in design and implementation of machine learning systems to augment clinical decision-making. To accomplish this goal, the candidate and mentors propose a systematic investigation of postoperative ‘patient acuity’ (i.e., risk for critical illness and death) and ‘intensity of care’ (i.e., triage destination and frequency of vital sign and laboratory measurements). After major surgery, misaligned patient acuity and intensity of care can lead to preventable harm and inappropriate resource use, affecting approximately 15 million inpatient surgeries annually in the US alone. When high-acuity patients receive low-intensity care, postoperative complications can progress to critical illness and cardiac arrest. Providing high-intensity care to low-acuity patients has low value and may cause harm through unnecessary treatments. It is difficult to address these problems systematically because there is no validated, unifying ‘intensity of care’ definition. The overall objective of this application is to understand intensity of care decision spaces in surgical patients and match them to clinical phenotypes and outcomes, leveraging this knowledge to generate precise, autonomous decision-support tools. The central hypothesis of this application is that inappropriate postoperative intensity of care is common, predictable, and associated with increased short- and long-term mortality, morbidity, and hospital costs. The rationale for this work is that integrating electronic health record data, machine learning, and clinical domain expertise offers opportunities to understand postoperative intensity of care decisions and develop decision-support tools capable of optimizing clinical outcomes and resource use. The specific aims of this proposal are to (1) develop and validate postoperative intensity of care definitions, (2) develop and validate interpretable, actionable acuity assessments that elucidate decision spaces, and (3) identify and predict postoperative intensity of care phenotypes. The proposed research is significant because it addresses a problem that affects millions of patients annually and is associated with potentially preventable harm and suboptimal resource use. The approach is innovative because the candidate and mentors are unaware of any prior attempts to classify and adjudicate postoperative intensity of care and understand the phenotypes and characteristics of patients receiving insufficient or excessive care. During the award period, the candidate will apply for an NIH-R01 investigator-initiated award for the prospective clinical implementation of an interpretable, actionable decision support tool incorporating validated intensity of care definitions and knowledge garnered from phenotype clustering, initially in a silent data collection period followed by a live period during which clinicians are provided with model outputs and clinically actionable recommendations.