背景:不同截断值的围手术期心肌损伤(PMI)与心脏手术后不同的预后效果相关。机器学习(ML)方法已广泛应用于心脏手术围手术期风险预测。然而,ML在PMI中的利用尚未研究。因此,我们试图开发和验证在体外循环(CPB)心脏手术中不同截断值PMI的ML表现.
方法:这是对多中心临床试验(OPTIMAL)的第二次分析,由于回顾性设计,放弃了书面知情同意的要求。2018年12月至2021年4月在中国招募18-70岁接受CPB择期心脏手术的患者。这些模型是使用阜外医院的数据开发的,并由其他三个心脏中心进行了外部验证。构建了传统逻辑回归(LR)和11个ML模型。主要结果是PMI,定义为术后最大心肌肌钙蛋白I超过参考上限的不同时间(40x,70x,100x,130x)我们通过检查接收器工作特性曲线(AUROC)下的面积来测量模型性能,精度-召回曲线(AUPRC),和校准布里尔分数。
结果:共有2983名符合条件的患者最终参与了模型开发(n=2420)和外部验证(n=563)。CatboostClassifier和RandomForestClassifier成为预测PMI的LR模型的潜在替代方法。AUROC显示四个截止值中的每一个都增加,在测试数据集中达到100xURL的峰值,在外部验证数据集中达到70xURL的峰值。然而,值得注意的是,AUPRC随着每个截止值的增加而下降。此外,Brier损失分数随着截止值的增加而减少,以130x的URL截止值达到最低点0.16。此外,CPB时间延长,主动脉持续时间,术前N端脑钠肽升高,术前中性粒细胞计数减少,较高的体重指数,高敏C反应蛋白水平的升高在所有4个临界值中被确定为PMI的危险因素.
结论:CatboostClassifier和RandomForestClassifer算法可以替代LR预测PMI。此外,术前较高的N末端脑钠肽和较低的高敏C反应蛋白是PMI的强危险因素,潜在机制需要进一步调查。
BACKGROUND: Perioperative myocardial injury (PMI) with different cut-off values has showed to be associated with different prognostic effect after cardiac surgery. Machine learning (ML) method has been widely used in perioperative risk predictions during cardiac surgery. However, the utilization of ML in PMI has not been studied yet. Therefore, we sought to develop and validate the performances of ML for PMI with different cut-off values in cardiac surgery with cardiopulmonary bypass (CPB).
METHODS: This was a second analysis of a multicenter clinical trial (OPTIMAL) and requirement for written informed consent was waived due to the retrospective design. Patients aged 18-70 undergoing elective cardiac surgery with CPB from December 2018 to April 2021 were enrolled in China. The models were developed using the data from Fuwai Hospital and externally validated by the other three cardiac centres. Traditional logistic regression (LR) and eleven ML models were constructed. The primary outcome was PMI, defined as the postoperative maximum cardiac Troponin I beyond different times of upper reference limit (40x, 70x, 100x, 130x) We measured the model performance by examining the area under the receiver operating characteristic curve (AUROC), precision-recall curve (AUPRC), and calibration brier score.
RESULTS: A total of 2983 eligible patients eventually participated in both the model development (n = 2420) and external validation (n = 563). The CatboostClassifier and RandomForestClassifier emerged as potential alternatives to the LR model for predicting PMI. The AUROC demonstrated an increase with each of the four cutoffs, peaking at 100x URL in the testing dataset and at 70x URL in the external validation dataset. However, it\'s worth noting that the AUPRC decreased with each cutoff increment. Additionally, the Brier loss score decreased as the cutoffs increased, reaching its lowest point at 0.16 with a 130x URL cutoff. Moreover, extended CPB time, aortic duration, elevated preoperative N-terminal brain sodium peptide, reduced preoperative neutrophil count, higher body mass index, and increased high-sensitivity C-reactive protein levels were identified as risk factors for PMI across all four cutoff values.
CONCLUSIONS: The CatboostClassifier and RandomForestClassifer algorithms could be an alternative for LR in prediction of PMI. Furthermore, preoperative higher N-terminal brain sodium peptide and lower high-sensitivity C-reactive protein were strong risk factor for PMI, the underlying mechanism require further investigation.