摘要
目的分析胰腺癌多器官转移(以下简称“PCMOM”)预后的危险因素,并基于机器学习算法构建PCMOM病人生存预测模型。方法从SEER数据库中提取2010~2015年间确诊为胰腺癌同时伴有多器官转移病人的临床资料,按照7∶3的比例将病人随机分为训练队列和内部验证队列。使用单因素联合多因素的比例风险回归模型(Cox模型)筛选出影响PCMOM病人预后的独立预后因素,利用筛选的独立预后因素构建Cox模型和随机生存森林模型,根据时间依赖性受试者操作特征曲线的表现选择Cox模型并对模型可视化。最后,使用夏普利算法来对变量权重进行排名,并使用校准曲线和决策曲线分析验证Cox模型的准确性和临床应用性。结果单因素联合多因素Cox模型分析结果显示,年龄、种族、转移部位、分化等级、肿瘤长径、手术和化疗是影响PCMOM预后的独立预后因素。在训练队列和内部验证队列的时间依赖性受试者操作特征曲线表现上Cox模型均优于随机生存森林模型。校准曲线和决策曲线分析分别说明Cox模型具有良好的预测准确性和临床应用性。夏普利算法对2种模型的可视化结果表明化疗、组织分级、转移部位和年龄对病人预后的影响排在前列。结论该研究构建的机器学习模型对于PCMOM病人的生存具有较好预测潜能,提示年龄≥50岁、其他种族、转移部位为肝+肺+骨、肿瘤长径≥50 mm、分化等级Ⅱ~Ⅳ级和其他是影响PCMOM病人预后的独立危险因素。转移部位为肝脏+骨或肺+骨、有过手术、有过化疗是影响PCMOM病人预后的独立保护因素。
Objective To identify risk factors for the prognosis of pancreatic cancer with multiple organ metastases(PCMOM)and to construct a survival prediction model based on machine learning algorithms to guide clinical practice.Methods Clinical data of PCMOM patients between 2010 and 2015 were extracted from the SEER(Surveillance,Epidemiology,and End Results)database.Patients were randomly divided into the training cohort and internal validation cohort at a 7∶3 ratio.Univariate and multivariate Cox regression analyses were used to screen for independent risk factors affecting the prognosis of PCMOM.A Cox regression model and a Random Survival Forest(RSF)通信作者:郑勇斌,Email:yongbinzheng@whu.edu.cn model were constructed using the identified independent risk factors.The Cox regression model was selected based on its performance on time-dependent receiver operating characteristic(ROC)curves(timeROC),followed by a visualization.Finally,the Shapley algorithm was used to rank variable importance,and calibration curves and decision curve analysis(DCA)were used to verify the accuracy and clinical applicability of the Cox regression model.Results Univariate and multivariate Cox regression analyses showed that age,race,site of metastasis,grade of differentiation,tumor diameter,surgery,and chemotherapy were independent factors affecting the prognosis of PCMOM.The Cox model outperformed the RSF model in timeROC performance in both the training and internal validation cohorts.The calibration curves and DCA indicated that the Cox regression model had good predictive accuracy and clinical applicability.Visualization results using the Shapley algorithm indicated that chemotherapy,histological grade,site of metastasis,and age were the most influential factors for the prognosis of PCMOM.Conclusion The machine learning model constructed in this study has good predictive potential for the survival of PCMOM patients,suggesting that age≥50 years,other races,metastatic sites involving liver+lung+bonetumor diameter≥50 mm,and differentiation gradeⅡ-Ⅳ,or other were identified as independent risk factors affecting the prognosis of patients with PCMOM.Metastatic sites involving liver+bone or lung+bone,a history of surgery,and a history of chemotherapy were identified as independent protective factors affectingthe prognosis of PCMOM patients.
作者
鲁聪
宋丹
王伟
王晨红
郑勇斌
Lu Cong;Song Dan;Wang Wei;Wang Chenhong;Zheng Yongbin(Department of Gastrointestinal Surgery,Renmin Hospital of Wuhan University,Hubei Wuhan 430060,China;Department of Hepatobiliary Surgery,East Hospital,Renmin Hospital of Wuhan University,Hubei Wuhan 430200,China)
出处
《腹部外科》
2025年第4期264-273,共10页
Journal of Abdominal Surgery
基金
中央高校基本科研业务费专项资金资助(2042022kf1100)。
关键词
机器学习
胰腺癌
多器官转移
预测模型
SEER数据库
Machine learning
Pancreatic cancer
Multi-organ metastasis
Prediction model
SEER database
