基于机器学习的狼疮肾炎患者肾小球微血栓形成诊断模型的建立  

Establishment of a diagnostic model for glomerular micro thrombosis in patients with lupus nephritis based on machine learning

作　　者：胡豪飞 许日聪 刘洋[1] 陈建宇 常哲毅 万启军[1] Hu Haofei;Xu Ricong;Liu Yang;Chen Jianyu;Chang Zheyi;Wan Qijun(Department of Nephrology,Shenzhen Second People's Hospital,the First Affiliated Hospital of Shenzhen University,Guangdong 518035,China)

机构地区：[1]广东省深圳市第二人民医院,深圳大学第一附属医院肾内科,深圳518035

出　　处：《中华风湿病学杂志》2022年第11期721-729,I0002,共10页Chinese Journal of Rheumatology

基　　金：国家自然科学基金(81900639);广东省深圳市科技创新委员会自然基金(JCYJ20190-80616-2807125);广东省深圳市医学重点学科建设项目(SZXK009)。

摘　　要：目的通过临床指标建立LN肾小球微血栓(GMT)形成的诊断模型。方法连续收集2010年1月至2021年3月在深圳市第二人民医院肾内科行肾脏穿刺活检诊断为LN患者。根据纳排标准对患者进行纳排。收集人口学资料、临床特点、生化指标及免疫指标。通过机器学习和Logistic逐步回归分析从上述候选指标中筛选出最重要变量,建立GMT联合诊断模型,并通过列线图呈现模型。受试者工作特征(ROC)曲线下的面积(AUC)、临床决策曲线和校准曲线分别用于评估模型区分度、临床使用价值和模型准确性。通过Bootstrap法重复采样500次对模型进行内部验证。结果符合纳排标准的共129例LN患者,女性117例(90.7%);平均年龄(34±11)岁。GMT患者39例(30.2%)。通过机器学习从47个备选变量筛选出重要性排在前10位的变量,然后通过Logistic逐步回归分析进一步筛选出5个变量用于建立GMT的诊断模型,分别为血红蛋白[OR(95%CI)=0.966(0.943,0.990),P=0.005]、血清补体C3[OR(95%CI)=0.133(0.022,0.819),P=0.030]、收缩压[OR(95%CI)=1.027(1.005,1.049),P=0.017]、淋巴细胞计数[OR(95%CI)=0.462(0.213,0.999),P=0.049]和凝血酶时间[OR(95%CI)=1.260(0.993,1.597),P=0.057]。得出LN患者GMT形成诊断模型的方程并建立列线图呈现模型。诊断模型[AUC(95%CI)=0.823(0.753,0.893)],说明模型具有较好的区分度。校准曲线分析结果提示模型预测GMT的风险与GMT实际发生风险相一致(Hosmer-Lemeshow检验,χ^(2)=14.62,P=0.067)。临床决策曲线和临床影响曲线反映模型具有较好的临床应用价值,尤其阈值概率在0.4~0.6临床应用价值更为显著。此外,经过Bootstrap法重复采样500次后,得出平均AUC(95%CI)=0.825(0.753,0.893),与原模型得出的AUC基本一致。结论基于临床指标通过机器学习和logistic逐步回归分析的方法建立了LN患者GMT形成的诊断模型。用于肾脏穿刺之前早期诊断GMT形成的风险。Objective To establish a diagnostic model for glomerular micro thrombosis(GMT)in lupus nephritis through clinical indicators.Methods A continuous collection of patients diagnosed with lupus nephritis(LN)by renal biopsy in the Department of Nephrology,Shenzhen Second People's Hospital,from January 2010 to March 2021.All patients were admitted and discharged through the inclusion and exclusion criteria.Demographic data,clinical characteristics,biochemical indicators,and immune indicators were collected.A GMT diagnosis model was established from the most important variables among the abovementioned variables through machine learning and Logistic stepwise regression analysis.The model was presented through a nomogram.The receiver operating characteristic curve(ROC),the clinical decision curve and the calibration curve were used to evaluate the model discrimination,clinical use and accuracy,respectively.The internal verification of the model was carried out by repeated sampling 500 times by the Bootstrap method.Results There were a total of 129 patients with lupus nephritis including the study,including 117 females(90.7%);the average age was(34±11)years.There were 39 patients with GMT(30.2%).Using machine learning to screen out the top 10 important variables from 47 candidate variables,then through logistic stepwise regression analysis,five variables were further screened to establish the diagnostic model of GMT,namely hemoglobin[OR(95%CI)=0.966(0.943,0.990),P=0.005],serum C3[OR(95%CI)=0.133(0.022,0.819),P=0.030],systolic blood pressure[OR(95%CI)=1.027(1.005,1.049),P=0.017],lymphocyte count[OR(95%CI)=0.462(0.213,0.999),P=0.049],and TT[OR(95%CI)=1.260(0.993,1.597),P=0.057].Draw up the equation of the GMT diagnosis model of lupus nephritis and establish a nomogram to present the model.The area under curve(AUC)of the diagnostic model was 0.823,95%CI(0.753,0.893),indicating that the model had a reasonable degree of discrimin-ation.The Hosmer-Lemeshow test showed a perfect fit between the predicted GMT risk and the observ

关 键 词：狼疮肾炎 肾小球 血栓形成 机器学习 Logistic逐步回归 

分 类 号：R593.242[医药卫生—内科学] TP181[医药卫生—临床医学]