Development And Application Of A Dynamic Prediction Model For Esophageal Cancer

Background:Prediction models are an indispensable part of the clinical practice,which is helpful to determine the best treatment strategy for individual patients.A drawback of the current available prediction model for esophageal cancer is that it can only be used to predict survival at the time of diagnosis or after treatment,ignoring the need to predict the 5-year survival at different predictive time points in the follow-up.In addition,the currently available prediction models assume that the impact of predictive variables on the overall survival rate of patients will not change over time,with a "time-constant effect".In fact,the hazard ratio of some variables may change during the follow-up period,which in turn has a dynamic impact on the 5-year survival rate of the patient.This effect is called a "time-varying effect." The purpose of this paper is to explore the variables with time-varying effect in patients with esophageal cancer,and to try to develop a dynamic prediction model,which can calculate the 5-year dynamic survival probability at different predictive time points during the follow-up period.Methods:The clinic-pathological information and survival data of 9132 patients with esophageal cancer obtained from SEER database were used to develop the prediction model.The covariates included age,marital status,sex,race,and primary location of tumor,histological type,pathological grade,T stage,N stage,M stage,surgery,chemotherapy and radiotherapy.Kaplan-Meier survival analysis was used to find out the prognostic factors of patients with esophageal cancer.The proportional baselines landmark supermodel(an extended COX model)is used to evaluate the time-varying effect of covariables and to develop a dynamic prediction model.The predictive model was externally validated using an independent Chinese patient cohort of 99 esophageal cancer patients.Including the validation of Discrimination and Calibration,such as C-index,AUC and the heuristic shrinkage factor.All the analyses were performed with R software(version 3.2.4),and the significant level was set to 0.05.Results:In Kaplan-Meier survival analysis,there was significant difference in survival among different groups of age,marital status,race,location of tumor,histological type,degree of differentiation,radiotherapy,operation,T stage,N stage and M stage(P<0.01).There was no significant difference in survival rate among different groups of chemotherapy(P = 0.803,0.058,respectively).In the various predictive variables of esophageal cancer patients,Age,location of primary tumor,histological type,chemotherapy,surgery and T stage showed significant time-varying effects on the overall survival rate.On the contrary,marital status,race,sex,differentiation,N stage and M stage showed a time-constant effect.The external validation results of the dynamic prediction model:C-index =0.746;AUC= 0.736,indicating that the prediction model has a good discrimination.The heuristic shrinkage factor= 0.996.The results showed that the model had good calibration.Conclusion:For esophageal cancer patients,the hazard ratio of age,prnmary tumor location,histological type,chemotherapy,and surgery and T stage will change over time,expressing a significant time-varying effect.The existence of time-varying effects indicates the importance of updating the 5-year survival probability during the follow-up period.In this study,the proportional baselines landmark supermodel in dynamic predictive analysis was used to predict the 5-year survival rate of patients with esophageal cancer at different time points for the first time.This new dynamic prediction model,which can update the patient's five-year survival probability over time,can be used to assist doctors in making better individualized treatment decisions based on dynamic assessment of the patient's prognosis.It can also be used to enhance patients' confidence and improve treatment compliance.