因果机器学习用于主动脉夹层个体化手术决策:全国性公平性临床AI框架
基于2016-2022年全国住院样本中45,112例主动脉夹层住院数据,开发整合预测、异质性治疗效应估计、公平性审计和个体化风险获益分析的因果推断框架。集成模型AUROC达0.845,CausalForestDML估计手术相关异质性效应,但所有平均治疗效应95%CI跨越零。框架支持风险分层和假设生成,不支持因果治疗推荐。
要点
- 集成模型预测院内死亡AUROC为0.845,校准因模型而异(Brier 0.076-0.161)。
- 手术相关条件平均治疗效应在≥80岁患者中39.5%为负,65-80岁为18.5%。
- 所有平均治疗效应95%CI跨越零,模型不能用于因果治疗推荐。
核心结论
- 该框架支持风险分层和关于治疗效应异质性的假设生成,但不支持因果治疗推荐。
- 临床部署需要前瞻性测试和独立外部验证。
研究设计与样本
- 使用2016-2022年HCUP全国住院样本出院数据,分析45,112例主动脉夹层住院。
- 五个ML模型加加权集成在2016-2019年数据(n=24,475)训练,在2020-2022年数据(n=20,624)时间验证。
- CausalForestDML估计手术相关异质性效应,并进行基于设定的稳健性分析。
- 审计种族/民族、保险状态和收入的公平性,并进行内部-外部地理验证。
主要发现
- 院内死亡率为11.0%(n=4,940)。
- 集成模型AUROC为0.845,COVID时期性能稳定。
- 校准因模型而异(Brier 0.076-0.161)。
- Causal forest估计手术相关异质性治疗效应(解释为观察性关联),≥80岁患者中39.5%显示负条件平均治疗效应,65-80岁为18.5%;但所有平均治疗效应95%CI跨越零。
- 夹层类型分层分析显示Type A(ATE +0.4 pp)与Type B(ATE +0.8 pp)模式不同。
临床怎么用
- 可用于主动脉夹层患者风险分层和生成治疗效应异质性假设,而非直接指导因果治疗决策。
- 临床使用前需前瞻性测试和独立外部验证。
局限性
- 所有平均治疗效应95%CI跨越零,不能得出因果治疗结论。
- 需要前瞻性测试和独立外部验证才能临床部署。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Most machine learning (ML) models for aortic dissection predict mortality but do not address treatment-effect heterogeneity, fairness, or bedside implementation. We developed a causal-inference-informed framework integrating prediction, heterogeneous treatment-effect estimation, fairness auditing, and individualized risk-benefit profiling. METHODS: Using 2016-2022 Healthcare Cost and Utilization Project National Inpatient Sample discharge data, we analyzed 45,112 aortic dissection hospitalizations. Five ML models, plus a weighted ensemble, were trained on 2016-2019 data (n = 24,475) and temporally tested on 2020-2022 data (n = 20,624). CausalForestDML estimated surgery-associated heterogeneous effects with specification-based robustness analyses. We audited fairness across race/ethnicity, insurance status, and income, and performed internal-external geographic validation. RESULTS: In-hospital mortality was 11.0% (n = 4,940). The ensemble achieved an area under the receiver operating characteristic curve (AUROC) of 0.845, with stable COVID-era performance. Calibration varied substantially across models (Brier 0.076-0.161). Causal forest estimated heterogeneous surgery-associated treatment effects (interpreted as observational associations), with 39.5% of patients aged ≥80 years showing negative conditional average treatment effects versus 18.5% aged 65-80; however, all average treatment effect (ATE) 95% confidence intervals crossed zero. Dissection type-stratified analysis showed distinct patterns for Type A (ATE +0.4 pp) versus Type B (ATE +0.8 pp). Fairness was consistent across the 3 largest racial groups (AUROC 0.820-0.862). Model discrimination was preserved with race excluded (AUROC 0.834 vs. 0.836). Geographic validation preserved discrimination (0.832 ± 0.007). CONCLUSION: This framework supports risk stratification and hypothesis generation regarding treatment-effect heterogeneity, but not causal treatment recommendations. Clinical deployment requires prospective testing and independent external validation.
引用
- Hassan AA, Elshami MI, Shaban Soliman AA, et al. Causal Machine Learning for Personalized Surgical Decision-Making in Aortic Dissection: A National Fairness-Aware Clinical AI Framework. Ann Vasc Surg. 2026. PMID: 41903841. doi:10.1016/j.avsg.2026.03.031
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。