2021 年第 5 期 第 16 卷

基于机器学习的血流动力学稳定的急性肺栓塞患者远期预后模型研究

Long-term prognostic model of hemodynamically stable patients with acute pulmonary embolism based on machine learning

作者：荣玮马珂于海旭刘卓慧李玉琳杜杰

英文作者：Rong Wei Ma Ke Yu Haixu Liu Zhuohui Li Yulin Du Jie

单位：首都医科大学附属北京安贞医院北京市心肺血管疾病研究所心血管生物研究室100029

英文单位：Cardiovascular Biology Laboratory Beijing Anzhen Hospital Capital Medical University Beijing Institute of Heart Lung and Blood Vessel Diseases Beijing 100029 China

关键词：急性肺栓塞；机器学习；远期预后；模型

英文关键词：Acutepulmonaryembolism;Machinelearning;Long-termprognosis;Model

• 摘要：

• 目的 应用机器学习构建并验证血流动力学稳定的急性肺栓塞患者远期预后模型。方法 收集2015年1月至2017年12月于首都医科大学附属北京安贞医院确诊的血流动力学稳定的155例急性肺栓塞患者的临床资料，电话随访患者出院后发生不良结局事件（死亡、再发静脉血栓栓塞症、大出血、心力衰竭）的情况。将所有患者按照6∶4比例随机分为训练集和验证集，采用最小绝对收缩和选择算子（LASSO）回归筛选用于构建模型的变量，采用随机森林算法在训练集中构建模型，采用受试者工作特征曲线评价模型在验证集中的预测效果。结果 155例患者中，23例（14.8%）患者发生远期不良结局事件。LASSO回归在训练集中筛选出12个预测因子用于构建模型（癌症、肺动脉高压、高同型半胱氨酸血症、贫血、深静脉血栓、心率、舒张压、静脉血栓栓塞症病史、4周内制动史、吸烟史、血清肌酐水平、右心室功能障碍）（回归系数分别为0.143、0.052、-0.023、-0.013、-4.325×10-9、-3.469×10-9、-2.924×10-9、-0.028、-3.904×10-9、0.102、0.037、-0.045），随机森林算法构建的模型在验证集中预测远期预后的受试者工作特征曲线下面积为0.77(95%置信区间：0.63～0.90)。结论 基于机器学习构建的模型对血流动力学稳定的急性肺栓塞患者发生远期不良结局具有一定的预测价值。

• Objective To construct and validate long-term prognostic model with machine learning in hemodynamically stable patients with acute pulmonary embolism(APE). Methods From January 2015 to December 2017, 155 hemodynamically stable patients with APE diagnosed in Beijing Anzhen Hospital, Capital Medical University were enrolled. The clinical data of paitents were collected, and the incidences of adverse outcome events were followed-up by phone(death, venous thromboembolism recurrence, major bleeding, and heart failure). The patients were randomly divided into a training set and a validation set according to the ratio of 6∶4. The least absolute shrinkage and selection operator(LASSO) regression was used to select the variables to construct the predictive model. Random forest algorithms was used to construct a predictive model in training set. The receiver operating characteristic curve was applied to evaluate the predictive effect of the model in validation set. Results Of 155 patients, 23 cases(14.8%) experienced the long-term adverse outcome events. Twelve predictors were screened out in training set by LASSO regression (cancer, pulmonary hypertension, hyperhomocysteinemia, anemia, deep vein thrombosis, heart rate, diastolic blood pressure, history of venous thromboembolism, history of immobility in past 4 weeks, history of smoking, serum creatinine level and right ventricular dysfunction)(the regression coefficients were as follows: 0.143, 0.052, -0.023, -0.013, -4.325×10-9, -3.469×10-9, -2.924×10-9, -0.028, -3.904×10-9, 0.102, 0.037, -0.045). The area under the receiver operating characteristic curve of the model constructed by random forest algorithms predicting long-term prognosis in validation set was 0.77(95% confidence interval: 0.63-0.90). Conclusion Long-term prognosis model based on machine learning has certain predictive value for the long-term adverse outcome in hemodynamically stable patients with APE.