预测乳腺癌术后死亡的两种模型建立及其性能和临床应用价值对比

Establishment of two models for predicting postoperative death of breast cancer and comparison of their performance and clinical application value

目的 比较随机生存森林(random survival forest,RSF)模型和Cox回归模型对乳腺癌患者术后死亡风险的预测性能。方法 回顾性分析徐州医科大学附属医院2014年1月至2016年12月收治的482例接受手术治疗的原发性乳腺癌患者的临床资料,按照患者随访期间生存情况,将其分别纳入存活组(n=446)和死亡组(n=36)。比较两组年龄、体重指数(body mass index,BMI)、临床分期、雌激素受体(estrogen receptor,ER)状态等资料,将Cox单因素回归分析中差异有统计学意义的因素分别使用RSF和Cox回归建立预测模型。使用受试者操作特征曲线(receiver operating characteristic curve,ROC曲线)评价两种模型预测乳腺癌患者术后死亡风险的曲线下面积(area under curve,AUC)和效能,并使用Kaplan-Meier曲线评价两种模型指导风险分层的价值,最后根据RSF模型预测结果建立在线动态预后预测分析网站。结果 随访截至2023年11月,所有患者均获得有效随访,中位随访时间93.30个月,随访期间死亡36例。多因素Cox回归显示,术前空腹血糖(fasting blood glucose,FBG)、术前尿酸(uric acid,UA)、N分期和ER表达与乳腺癌患者术后死亡风险均显著相关(均P<0.05)。RSF最优模型包括年龄、N分期、病理分级、术前UA、术前癌胚抗原(carcinoembryonic antigen,CEA)、肿瘤最大径、术前FBG、术前糖类抗原153、放疗和化疗方案这10个变量。RSF模型1、3、5年AUC、敏感度和特异度均高于Cox回归模型,Brier分数均低于Cox模型,整体预测性能显著优于Cox模型(P<0.05)。对患者风险分层后,两种模型低风险组和高风险组生存比较差异均有统计学意义(均P<0.05)。在线动态预后预测分析网站操作方便,临床医生只需填入相关信息,即可直接得到患者预后生存情况。结论 基于RSF和Cox回归建立的模型均能够为乳腺癌患者术后死亡风险的预测提供可靠参考,但RSF模型的预测性能和稳定性略优于Cox模型,结合在线实时动态预测患者总体生存情况和不同时间点的死亡风险情况,具有更高的临床指导意义和价值。

Objective To compare the predictive performance of random survival forest(RSF)and Cox regression on postoperative mortality risk of breast cancer patients.Method The clinical data of 482 patients with primary breast cancer treated by surgery in the Affiliated Hospital of Xuzhou Medical University from January 2014 to December 2016 were retrospectively analyzed.According to their survival during the followup period,they were divided into survival group(n=446)and death group(n=36)respectively.The data such as age,body mass index(BMI),clinical stage and estrogen receptor(ER)status were compared between the two groups.The factors with statistical differences between the two groups were established by RSF and Cox regression.Receiver operating characteristic curve(ROC curve)was used to evaluate the area under the curve(AUC)and efficacy of the two models in predicting the postoperative mortality risk of breast cancer patients,and KaplanMeier curve was used to evaluate the value of the two models in guiding risk stratification.Finally,an online dynamic prognostic prediction analysis website was developed based on the RSF model.Result All patients were followed up effectively until November,2023.The median followup time was 93.30 months,and 36 patients died during the followup.Multivariate Cox regression showed that preoperative fasting blood glucose(FBG),preoperative uric acid(UA),axillary lymph node metastasis and ER expression were related to postoperative death risk of breast cancer patients(all P<0.05).The optimal model of RSF included ten variables:namely,age,N stage,pathological grade,preoperative UA,maximum tumor diameter,preoperative FBG,preoperative glycoprotein 153,preoperative carcinoembryonic antigen(CEA),radiation therapy,and chemotherapy regimen.In the dataset,the AUC,sensitivity,specificity and of RSF model at 1 year,3 years and 5 years were significant higher than that of Cox regression model(all P<0.05),while Brier scores were significant lower than that of Cox regression model,so the overall prediction performance was significant better than that of Cox regression model(P<0.05).After stratification of patients'risk,the survival difference between the lowrisk group and the highrisk group of the two models were statistically significant(all P<0.05).The online dynamic prognostic prediction analysis website was easy to operate.Clinicians could directly obtain the prognosis and survival of patients byfilling in relevant information.Conclusion The models based on RSF and Cox regression can provide reliable reference for predicting the risk of postoperative death of breast cancer patients,but the predictive performance and stability of RSF model are slightly better than Cox model,which has higher clinical significance and value combined with realtime online dynamic prediction of overall patient survival and the risk of death at different time points.