作者：胡文标，刘礼敬，林晓峰，宋清华，陆元喜，韦佳玲

单位：

广西壮族自治区南宁市第二人民医院心血管内科

单位（英文）：

Department of Cardiovascular Medicine, the Second Nanning People's Hospital, Nanning 530031, China

关键词：

心肌梗死; 心力衰竭; 决策树; 梯度提升决策树; Logistic模型; 预测;

关键词（英文）：

Myocardial infarction; Heart failure; Decision trees; Gradient boosted decision trees; Logistic models; Forecasting

中图分类号：

R 542.22 R 541.62

DOI：

10.12114/j.issn.1008-5971.2023.00.133

基金项目：

摘要：

目的 探讨梯度提升决策树（GBDT）在预测急性心肌梗死（AMI）患者PCI后住院期间发生心力衰竭（HF）中的应用价值。方法 回顾性选取2021—2022年于南宁市第二人民医院行PCI的AMI患者200例为研究对象。将患者分为训练集（145例）和测试集（55例）。根据PCI后住院期间HF发生情况，将训练集患者分为HF组（48例）和非HF组（97例）。收集患者一般资料及PCI前实验室检查指标、心脏彩超检查指标。基于单因素分析结果，采用R4.1.2软件分别构建预测AMI患者PCI后住院期间HF发生风险的GBDT算法模型和Logistic回归模型；分别采用ROC曲线、校准曲线分析GBDT算法模型、Logistic回归模型的区分度、准确性。结果 HF组年龄大于非HF组，有糖尿病病史者占比、超敏C反应蛋白（hs-CRP）、白细胞计数（WBC）、中性粒细胞计数、肌酸激酶同工酶（CK-MB）高于非HF组（P<0.05）。将单因素分析中差异有统计学意义的指标纳入GBDT算法模型，通过GBDT算法获得这6项指标的相对重要性，由小到大依次为糖尿病病史（2.220）、中性粒细胞计数（7.713）、年龄（14.734）、CK-MB(16.819)、WBC(24.828)、hs-CRP(33.686)。多因素Logistic回归分析结果显示，年龄、hs-CRP、WBC、中性粒细胞计数、CK-MB是训练集AMI患者PCI后住院期间发生HF的影响因素（P<0.05），构建Logistic回归模型，其具体公式为：logit(P)=-18.182+0.147×年龄+0.233×hs-CRP+0.438×WBC+0.242×中性粒细胞计数+0.003×CK-MB。ROC曲线分析结果显示，GBDT算法模型、Logistic回归模型预测训练集AMI患者PCI后住院期间发生HF的AUC分别为0.989[95%CI(0.974,1.000)]、0.864[95%CI(0.786,0.942)]；GBDT算法模型、Logistic回归模型预测测试集AMI患者PCI后住院期间发生HF的AUC分别为0.900[95%CI(0.817,0.982)]、0.763[95%CI(0.639,0.888)]。校准曲线分析结果显示，GBDT算法模型、Logistic回归模型预测训练集、测试集AMI患者PCI后住院期间发生HF的概率分别与本组AMI患者PCI后住院期间HF的实际发生率一致。结论 本研究基于年龄、糖尿病病史、hs-CRP、WBC、中性粒细胞计数和CK-MB 6个指标构建的GBDT算法模型对AMI患者PCI后住院期间发生HF有较好的预测价值，且优于传统Logistic回归模型，这可为AMI患者PCI后预后的评估及干预治疗提供参考依据。

英文摘要：

Objective To explore the application value of gradient boosted decision trees (GBDT) in predicting heart failure (HF) in patients with acute myocardial infarction (AMI) during hospitalization after PCI. Methods A total of 200 patients with AMI who underwent PCI in the Second Nanning People's Hospital from 2021 to 2022 were retrospectively selected as the study subjects. The patients were divided into a training set (145 cases) and a test set (55 cases) . According to the occurrence of HF during hospitalization after PCI, the patients in training set were divided into the HF group (48 cases) and the non-HF group (97 cases) . The general data, laboratory examination indexes and cardiac ultrasound examination indicators before PCI of patients were collected. Based on the results of univariate analysis, GBDT algorithm model and Logistic regression model were constructed by R 4.1.2 software to predict the risk of HF in AMI patients during hospitalization after PCI. ROC curve and calibration curve were used to analyze the discrimination and accuracy of GBDT algorithm model and Logistic regression model. Results The age of the HF group was older than that of the non-HF group, the proportion of patients with a history of diabetes, high-sensitivity C-reactive protein (hs-CRP) , white blood cell count (WBC) , neutrophil count, and creatine kinase isoenzyme (CK-MB) were higher than those in the non-HF group (P < 0.05) . The indicators with statistically significant differences in univariate analysis were included in the GBDT algorithm model, and the relative importance of these six indicators was obtained through the GBDT algorithm, and the descending order was diabetes history (2.220) , neutrophil count (7.713) , age (14.734) , CK-MB (16.819) , WBC (24.828) , hs-CRP (33.686) . Multivariate Logistic regression analysis showed that age, hs-CRP, WBC, neutrophil count and CK-MB were the influencing factors of HF in AMI patients during hospitalization after PCI in the training set (P < 0.05) . A Logistic regression model was constructed, with the specific formula as follows: logit (P) =-18.182+0.147×age+0.233×hs-CRP+0.438×WBC+0.242×neutrophil count+0.003×CK-MB. The results of the ROC curve analysis showed that the AUC of the GBDT algorithm model and Logistic regression model for predicting HF in AMI patients during hospitalization after PCI in the training set was 0.989 [95%CI (0.974, 1.000) ] and 0.864 [95%CI (0.786, 0.942) ] . The AUC of the GBDT algorithm model and Logistic regression model for predicting HF in AMI patients during hospitalization after PCI in the test set was 0.900 [95%CI (0.817, 0.982) ] and 0.763 [95%CI (0.639, 0.888) ] . The results of the calibration curve analysis showed that the probability of HF in AMI patients during hospitalization after PCI in the training set and test set predicted by the GBDT algorithm model and Logistic regression model was consistent with the actual incidence of HF in AMI patients during hospitalization after PCI, respectively. Conclusion In this study, the GBDT algorithm model constructed based on age, diabetes history, hs-CRP, WBC, neutrophil count and CK-MB has an excellent predictive value for HF in AMI patients during hospitalization after PCI and is superior to the traditional Logistic regression model, which can provide a reference for the evaluation of prognosis and intervention treatment of AMI patients after PCI.

参考文献：