Application Of Cox Model Based On Neural Network In Survival Analysis Of Patients With Chronic Heart Failure

Objective:In the past,the Cox proportional hazards model was mostly used to assess the prognosis of patients with heart failure,but it was limited by the assumption of linear proportional hazards.In this study,we built a Cox model based on neural network to remove the linear and proportional risk limitations of the classic Cox model,and modeled the data of patients with chronic heart failure with complex relationships between variables to achieve a more accurate individualized prognostic risk assessment.Identify high-risk patients and guide clinicians to choose appropriate treatment options to improve survival and avoid unnecessary treatments.Methods:This study contains two parts of data:1.Four commonly used 4 benchmark databases in the field of survival analysis（SUPPORT,METABRIC,GBSG,FLCHAIN）;2.Collection of two top three hospitals in Shanxi Province from January 1,2014 to 2019During February 15th,the medical records of patients who were clearly diagnosed with chronic heart failure and met the criteria for inclusion in this study.Firstly,four benchmark databases with different numbers of variables,different censorship rates,and ratio winds are assumed to meet different conditions to verify the distinction and calibration capabilities of the traditional Cox model and the three neural network-based Cox models（Deepsurv,Cox-CC,Cox-Time）.Then,in the data of patients with heart failure,the variables were initially screened by the single-factor Cox model,and the Cox,Lasso-Cox and three neural network-based Cox models（Deepsurv,Cox-CC,Cox-Time）were constructed respectively,and the performance of different models was discussed using C^tdand IBS.Results:On the four benchmark data sets,in the performance of C^tdand IBS two evaluation indicators,the classic Cox is usually the worst than the performance.The C^tdand IBS results of Deep Surv and Cox-CC,which have similar basic principles,are very close to each other.The overall performance of Cox-Time is the best,achieving the best C^tdand IBS on the three data sets.However,the performance of this variable basically meets the risk ratio assumption on the data set is average.But this variables basically satisfy the proportional hazard assumption on FLCHAIN,the performance is average.2.In the data of heart failure patients,the performance of the Cox model in C^tdand IBS is the worst among all models(C^td:0.776,IBS:0.098).The performance of Lasso-Cox is better than that of Cox model(C^td:0.794,IBS:0.096).The performances of the two models,Deepsurv and Cox-CC,which take into account the nonlinear relationship of the covariates,are not much different.The C^tdis 0.816 and 0.812,respectively,and the IBS is 0.093,and 0.094,respectively,and both have good discrimination and calibration capabilities.The C^tdof the Cox-Time model is 0.829 and the IBS is 0.084.The two indicators are the best performance in all models,indicating that it is more accurate in predicting the prognosis of patients and can more accurately assess the prognostic survival rate and survival time of patients.Conclusion:In this study,three Cox model based on neural network were constructed to assess the prognosis risk of patients with chronic heart failure,and compared them with the traditional Cox model and Lasso-Cox model.The results showed that the Deep Surv and Cox-CC models used neural network characteristics to well described the complex non-linear relationship between variables and achieved a more traditional The model performs better.The Cox-Time model further expands the proportional hazard assumptions on this basis,and had better distinguishing and calibration capabilities than the Deep Surv and Cox-CC models.Cox-Time,a less restrictive model,is more suitable for real data that has complex relationships between variables and does not meet certain specific assumptions.The Cox model based on neural network can improve the accuracy of prediction,carry out individualized risk assessment of patients with heart failure,identify high-risk patients,and provide a decision-making basis for effective intervention of patients,which has practical application value.