腹膜后肉瘤预后模型的构建与验证:一项基于机器学习算法的大样本真实世界研究  

Development and validation of survival predicting model in patients with retroperitoneal sarcoma:A large real-world study based on machine learning

作　　者：白斗[1] 黄坤 Bai Dou;Huang Kun(Department of General Surgery(Vascular Surgery),Mianyang Central Hospital,Mianyang 621099,China;Department of General Surgery,Mianyang Hospital of Traditional Chinese Medicine,Mianyang 621000,China)

机构地区：[1]绵阳市中心医院普外(血管)外科,四川绵阳621099 [2]绵阳市中医医院普外科,四川绵阳621000

出　　处：《中国血管外科杂志（电子版）》2024年第2期172-179,共8页Chinese Journal of Vascular Surgery(Electronic Version)

摘　　要：目的构建精准化、个体化评价腹膜后肉瘤(retroperitoneal sarcoma,RS)患者预后的模型,为临床决策制定提供参考。方法提取监测、流行病学及结局(Surveillance,Epidemiology and End Results,SEER)数据库2000~2019年经病理学确诊的RS患者临床资料,以7∶3的比例随机划分为训练集和验证集,采用多因素Cox比例风险模型、LASSO回归模型和随机生存森林模型筛选变量,构建预测1、3年肿瘤特异性生存期(cancer-specific survival,CSS)和总生存期(overall survival,OS)的诺谟图模型,利用一致性指数、受试者工作特征曲线评估其预测价值,并用校正曲线对诺谟图预测模型进行内部(训练集)与外部(验证集)验证。结果共纳入2559例患者,其中训练集1791例,验证集768例。多因素Cox比例风险模型显示,年龄、肿瘤分期、肿瘤分级、手术方式和化疗是OS的独立影响因素,而年龄、病理类型、肿瘤分级、肿瘤分期、手术方式和化疗是CSS的独立影响因素。LASSO回归模型显示,年龄、性别、病理类型、肿瘤分级、肿瘤分期、手术方式和化疗患者OS相关,而年龄、病理类型、肿瘤分级、肿瘤分期、手术方式和化疗与患者CSS相关。随机生存森林模型显示,影响OS的重要性评分前5位变量分别为肿瘤分级、手术方式、年龄、肿瘤分期和化疗,而影响CSS的重要性评分前5位变量分别为肿瘤分级、手术方式、肿瘤分期、化疗和年龄。基于上述因素所构建的诺谟图验证结果表明,OS在训练集和验证集的一致性指数分别为0.746(95%CI=0.730~0.760)和0.729(95%CI=0.710~0.750),而CSS分别为0.770(95%CI=0.750~0.790)和0.743(95%CI=0.720~0.770),校准曲线表现出良好的一致性。结论年龄、性别、肿瘤分级、肿瘤分期、手术和化疗是RS患者预后的独立影响因素。本研究构建的诺谟图预测模型具有良好的预测价值,有利于临床对RS患者选择个性化治疗。Objective To construct a accurate and individual survival predicting model for retroperitoneal sarcoma(RS)patients using machine learning,in order to provide reference for clinical decision-making.Methods The clinical data of RS patients diagnosed via pathology from 2000 to 2019 were extracted from Surveillance,Epidemiology and End Results(SEER)database.The patients were randomly divided into a training set and a verification set in a 7∶3 ratio.Multivariate Cox proportional hazard model,LASSO regression and Random Survival Forest model were used to screen independent prognostic variables used to construct a Nomogram model for predicting tumor specific survival(CSS)and total survival(OS)at 1-and 3-year,respectively.The predictive value was evaluated by C-index and receiver operating characteristic curve,and the calibration curve used to verify the Nomogram prediction model internally(training set)and externally(verification set).Results A total of 2559 patients were included,including 1791 in the training set and 768 in the verification set.Multivariate Cox proportional hazard model analysis showed that the age,tumor stage,tumor grade,surgery and chemotherapy were independent influencing factors for OS,and age,histologic patterns of tumor,tumor stage,tumor grade,surgery and chemotherapy were independent influencing factors for CSS.LASSO regression analysis showed that the age,gender,histologic patterns,tumor stage,tumor grade,surgery and chemotherapy were associated with OS,and age,histologic patterns,tumor stage,tumor grade,surgery and chemotherapy were associated with CSS.For OS,the top five variables for importance scores via random Survival Forest model were the tumor grade,surgery,age,tumor stage and chemotherapy,and for CSS,were the tumor grade,surgery,tumor stage,chemotherapy and age.The Nomogram constructed based on these factors show that the C-index of OS in training set and verification set were 0.746(95%CI=0.730-0.760)and 0.729(95%CI=0.710-0.750),respectively,and the C-index of CSS were 0.770(95%CI=0

关 键 词：腹膜后肉瘤 预后 预测模型 生存分析 机器学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] R735.4[自动化与计算机技术—控制科学与工程]