Modern Analytics to Improve Quality & Outcome Assessments Following Congenital Heart Surgery

先天性心脏缺陷（CHD）是最常见的出生缺陷，每100名新生儿中就有1名患有CHD。而 随着时间的推移，手术结果有所改善，某些病变的发病率和死亡率仍然很高， 在大约120个美国先天性心脏病中心中存在显著差异。这加强了对 先天性心脏病项目，有几个国家的高调报告导致项目关闭， 提高质量的举措，以改善国家成果。所有这些举措都依赖于严格的 程序质量的措施，但如何最好地衡量和报告质量仍然存在争议， 替补演员CHD数据的几个特征和当前的分析方法使质量评估复杂化。第一章 CHD是一种广泛异质性的、不同病变的稀疏混合。目前的方法结合了联合收割机的所有CHD 仅根据手术类型和少数一般情况对手术进行质量报告和风险调整 尽管其他常规收集的变量的重要影响。机器学习方法是 能够更好地结合这些高维临床数据，但仍然研究不足， 依赖过时的回归。2)初步研究表明，治疗的冠心病类型存在很大差异 然而，医院病例组合的差异尚未量化，也未用于质量 报告.评估医院病例组合分布的方法可以更好地理解 哪些医院治疗可比患者，哪些患者子集足够大， 质量评估，但尚未研究。3)目前的方法仅支持CHD的间接评估 程序质量，禁止直接比较利益和忽视不确定性的重要来源，导致 质量评估的不准确性。因果推理框架可以支持更直接和临床的 医院之间的相关质量比较，但尚未进行调查。根据NOT-HL-19-712， 我们建议采用当代统计方法来应对这些挑战。目标1将优化风险- 使用机器学习方法进行调整，以表征手术发病率和死亡率。目标2将 识别医院间重叠和独特的CHD患者亚组，以改善医院特征 病例组合并进行有效的医院比较。目标3将制定稳健的医院质量评估， 使用因果推理方法支持医院之间直接和临床相关的质量比较， 重点放在重叠的领域，并包括医院随机效应，以解释目前被忽视的 不确定性我们的方法将应用于STS先天性心脏手术数据库， >90%的美国先天性心脏病医院的临床信息，并在外部来源，儿科 心脏重症监护联盟（PC 4）登记研究。这些进展将支持国家倡议， 改善受CHD影响的儿童和家庭的质量和结果，还将提供一个框架， 可以广泛应用于改善其他领域的质量评估。

Congenital heart defects (CHD) are the most common birth defects, impacting ~1 in every 100 births. While surgical outcomes have improved over time, morbidity and mortality remain high for certain lesions, and significant variation exists across the ~120 U.S. congenital heart centers. This has heightened the scrutiny of congenital heart programs, with several national high-profile reports leading to program closure and several quality improvement initiatives to improve national outcomes. All such initiatives depend upon rigorous measures of program quality, but how best to measure and report quality remains controversial and understudied. Several features of CHD data and current analytic methods complicate quality assessments. 1) CHD is a widely heterogeneous, sparse mix of different lesions. Current approaches combine all CHD surgeries together for quality reporting and adjust for risk based only on operation type and few general factors, despite the important influence of other routinely collected variables. Machine learning approaches are better able to incorporate these high-dimensional clinical data but remain understudied, and current methods rely upon outdated regression. 2) Preliminary studies suggest wide variability in the types of CHD treated across hospitals, yet differences in hospital case-mix have not been quantified and are not used in quality reporting. Approaches to assess hospital case-mix distributions could provide an improved understanding of which hospitals treat comparable patients and which patient subsets are large enough to allow meaningful quality assessments but have not been studied. 3) Current methods support only indirect assessments of CHD program quality, prohibiting direct comparisons of interest and ignore important sources of uncertainty, leading to inaccuracies in the quality estimates. A causal inference framework could support more direct and clinically relevant quality comparisons across hospitals but has not been investigated. In response to NOT-HL-19-712, we propose to apply contemporary statistical approaches to address these challenges. Aim 1 will optimize risk- adjustment using machine learning approaches to characterize operative morbidity and mortality. Aim 2 will identify overlapping and unique CHD patient subgroups across hospitals to improve characterization of hospital case-mix and enable valid hospital comparisons. Aim 3 will develop robust hospital quality estimates that will support direct and clinically relevant quality comparisons across hospitals using causal inference approaches, focusing on areas of overlap, and including hospital random effects to account for currently ignored sources of uncertainty. Our methods will be applied within the STS Congenital Heart Surgery Database which captures clinical information >90% of U.S. congenital heart hospitals and validated in an external source, the Pediatric Cardiac Critical Care Consortium (PC4) Registry. These advancements will support national initiatives to improve quality and outcomes for children and families impacted in CHD and will also provide a framework that can be applied broadly to improve quality assessment across other fields.