【摘要】 目的 探讨机器学习预测微创心脏外科手术（minimally invasive cardiac surgery, MICS）患者术中红细胞输注概率的可行性。方法 选取2019年6月至2022年12月广东省人民医院行MICS的患者1 837例，按照8.5∶1.5的比例随机分层划分为训练集（1 562例）和测试集（275例）。使用训练集在H2O.ai平台自动进行模型训练，然后利用测试集对模型进行内部验证以选择最佳模型，并对该模型进行可视化解释。结果 1 837例MICS患者中，男820例，女1 017例，年龄18～83岁，平均（48.3±15.0）岁。MICS术中输注红细胞的发生率为11.3%(207/1 837)。自动化建立的59个机器学习模型中，基于梯度提升机（gradient boosting machine, GBM）算法的模型预测性能最佳（AUC=0.880,95%CI:0.832～0.927,P=0.029）。GBM模型中最关键的变量是体外循环（cardiopulmonary bypass, CPB）最低红细胞压积（hematocrit,HCT），其权重百分比为23.54%，其次是术前HCT(16.01%)、CPB时间（9.09%）、外科医生团队（8.36%）、BMI(4.94%)、身高（4.47%）、心脏停搏液类型（4.45%）等。结论 在建立的机器学习模型中，基于GBM算法的模型表现出最佳性能，可以有效、准确地预测红细胞输注风险，提供了一个可解释的预测模型，有助于临床医生推进MICS患者的科学用血。更多还原

【Abstract】 Objective To explore the feasibility of machine learning in predicting the probability of intraoperative red blood cell transfusion in patients undergoing minimally invasive cardiac surgery(MICS). Methods A total of 1 837 patients who underwent MICS in Guangdong Provincial People’s Hospital from June 2019 to December 2022 were selected, and were randomly divided into training sets(1 562 cases) and testing sets(275 cases) in an 8.5:1.5 ratio. The training set was used for automated model training on the H2O.ai platform. And the testing set was used for internal validation to select the best model,which was further visually interpreted. Results Among 1 837 MICS patients, there were 820 males and 1 017 females, aged from 18 to 83 years, with an average age of(48.3 ± 15.0) years. The incidence of intraoperative red blood cell transfusion in MICS was 11.3%(207/1 837). Among the 59 machine learning models developed through automation, the model based on the gradient boosting machine(GBM) algorithm exhibited the best predictive performance(AUC = 0.880, 95% CI: 0.832-0.927, P = 0.029). The most critical variables in the GBM model were the nadir hematocrit(HCT) during cardiopulmonary bypass(CPB), with a weight percentage of 23.54%, followed by preoperative HCT(16.01%), CPB duration(9.09%), surgeon team(8.36%), BMI(4.94%), height(4.47%), and type of cardioplegia solution(4.45%). Conclusions In the established machine learning models, the GBM model outperformed other machine learning models in predicting intraoperative red blood cell transfusion in MICS. This innovative approach provides an interpretable predictive model that can assist clinical practitioners in evidence-based transfusion practices for patients undergoing MICS.更多还原