肺切除术前心肺运动试验：仍有指征？基于机器学习的预测效用评估
该研究基于两项前瞻性多中心研究的二次分析，纳入497例肺切除候选者，采用五种机器学习模型评估CPET对术后肺部并发症（PPCs）和心血管并发症（PCCs）的增量预测价值。结果显示，在未选择患者及符合ACCP或ERS/ESTS指南指征的亚组中，CPET参数均未改善模型预测性能（PPCs的AUC-ROC=0.72-0.78，p=0.47；PCCs的AUC-ROC=0.65-0.73，p=0.96）。结论提示当代外科实践中CPET的术前风险分层作用需重新评估。
要点
- 在497例肺切除患者中，CPET未提升机器学习模型对PPCs或PCCs的预测能力。
- 符合ACCP或ERS/ESTS指南指征的亚组中，CPET同样未改善预测性能。
- 现行指南中CPET的术前风险分层指征可能需要重新审视。
核心结论
- 在当代外科实践中，CPET未改善机器学习模型对肺切除术后肺部并发症（PPCs）或心血管并发症（PCCs）的预测性能，无论患者是否符合既定指南指征。
- 肺切除术前风险分层中CPET的角色应重新评估。
研究设计与样本
- 研究为两项前瞻性多中心研究的二次分析，纳入497例适合CPET且接受术前肺功能测试（PFTs）和动脉血气分析的肺切除候选者。
- 训练了五种机器学习模型，采用嵌套交叉验证预测并发症，基于包括AUC-ROC在内的四项指标比较预测性能。
- 术后肺部并发症（PPCs）和心血管并发症（PCCs）在住院期间记录。
主要发现
- PPCs发生于71例（14%）患者；将CPET参数加入PFTs和基线临床数据后，未改善未选择患者（AUC-ROC=0.72-0.78；p=0.47）、符合ACCP标准者（n=236；AUC-ROC=0.64-0.78；p=0.70）或符合ERS/ESTS标准者（n=168；AUC-ROC=0.59-0.76；p=0.92）的PPCs预测能力。
- PCCs发生于90例（18%）患者；CPET参数同样未改善未选择患者（AUC-ROC=0.65-0.73；p=0.96）、ACCP亚组（AUC-ROC=0.61-0.73；p=0.82）或ERS/ESTS亚组（AUC-ROC=0.62-0.69；p=0.87）的PCCs预测性能。
关键数值与推荐
- PPCs预测的AUC-ROC范围：未选择患者0.72-0.78，ACCP亚组0.64-0.78，ERS/ESTS亚组0.59-0.76。
- PCCs预测的AUC-ROC范围：未选择患者0.65-0.73，ACCP亚组0.61-0.73，ERS/ESTS亚组0.62-0.69。
临床怎么用
- 对于肺切除候选者，基于PFTs和基线临床数据的机器学习模型已具备一定预测能力，加入CPET参数未带来额外获益。
- 在符合ACCP或ERS/ESTS指南CPET指征的患者中，CPET的常规使用可能需重新考量。
局限性
- 本研究为二次分析，结果需在前瞻性研究中验证。
- 未报告CPET具体参数、模型类型细节及外部验证结果。
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
RATIONALE: Despite significant advances in patient care and outcomes, criteria for cardiopulmonary exercise testing (CPET) in risk stratification guidelines for lung resection have not been updated in over a decade. We hypothesised that CPET no longer holds additional predictive value for postoperative complications. METHODS: In this secondary analysis, we included lung resection candidates from two prospective, multicentre studies eligible for CPET and assessed with preoperative pulmonary function tests (PFTs) and arterial blood gas analysis. Postoperative pulmonary (PPCs) and cardiovascular complications (PCCs) were documented during hospitalisation. We trained five types of machine learning models applying nested cross-validation to predict complications and compared predictive performance based on four metrics, including area under the receiver operating characteristic curve (AUC-ROC). RESULTS: A total of 497 patients were included. PPCs developed in 71 (14%) patients. Adding CPET parameters to PFTs and baseline clinical data did not improve the ability of models to predict PPCs in unselected patients (AUC-ROC=0.72-0.78; p=0.47), nor in those meeting American College of Chest Physicians (ACCPs) (n=236; AUC-ROC=0.64-0.78; p=0.70) or European Respiratory Society/European Society of Thoracic Surgery (ERS/ESTS) criteria (n=168; AUC-ROC=0.59-0.76; p=0.92). PCCs developed in 90 (18%) patients. CPET parameters likewise did not improve model performance for the prediction of PCCs in unselected patients (AUC-ROC=0.65-0.73; p=0.96), nor in the ACCP (AUC-ROC=0.61-0.73; p=0.82) or ERS/ESTS subgroups (AUC-ROC=0.62-0.69; p=0.87). CONCLUSIONS: In contemporary surgical practice, CPET did not improve the predictive performance of machine learning models for PPCs or PCCs in patients with an indication based on established guidelines or in those without. The role of CPET in preoperative risk stratification for lung resection should be re-evaluated. TRIAL REGISTRATION NUMBER: NCT03498352, NCT04826575.
引用
- Filakovszky Á, Brat K, Tschoellitsch T, et al. Cardiopulmonary exercise testing before lung resection surgery: still indicated? Evaluating predictive utility using machine learning. Thorax. 2026. PMID: 41043965. doi:10.1136/thorax-2024-221485
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。