Ishwaran Hemant, Blackstone Eugene H
Division of Biostatistics, Miller School of Medicine, University of Miami, Miami, USA.
Department of Quantitative Health Sciences, Cleveland Clinic, Cleveland, OH, USA.
Comput Struct Biotechnol J. 2025 Aug 27;28:312-320. doi: 10.1016/j.csbj.2025.08.017. eCollection 2025.
Traditional methods for evaluating hospital performance, such as regression or propensity score analysis, offer population-level comparisons but lack the granularity required for patient-level insight. We propose a causal framework based on virtual (digital) twins, enabling counterfactual outcome comparisons for individual patients across hospitals. Using data from the American Association for Thoracic Surgery (AATS) Quality Gateway Adult Cardiac Database, which includes 52,792 surgeries across 19 hospitals, we estimate patient-level causal effects for adverse surgical outcomes. Our approach combines model-free variable priority screening, random forests quantile classification (RFQ) for handling rare events, and isolation forests to assess treatment overlap and exclude invalid counterfactuals. Building on prior work, we introduce graphical tools for overlap diagnostics and counterfactual visualization at both the institutional and patient level. These tools reframe outcome modeling as individualized causal inference and support transparent, patient-centered hospital benchmarking.
评估医院绩效的传统方法，如回归分析或倾向得分分析，可提供总体层面的比较，但缺乏深入了解患者层面情况所需的精细度。我们提出了一个基于虚拟（数字）双胞胎的因果框架，能够对各医院的个体患者进行反事实结果比较。利用美国胸外科医师协会（AATS）质量网关成人心脏数据库的数据，该数据库包含19家医院的52792例手术，我们估计了不良手术结果的患者层面因果效应。我们的方法结合了无模型变量优先级筛选、用于处理罕见事件的随机森林分位数分类（RFQ）以及隔离森林，以评估治疗重叠并排除无效的反事实情况。在先前工作的基础上，我们引入了图形工具，用于机构和患者层面的重叠诊断和反事实可视化。这些工具将结果建模重新构建为个性化因果推断，并支持透明的、以患者为中心的医院基准评估。
