Joint Modeling Of Longitudinal And Survival Data With A Random Forest Based Association Structure

In recent years,jointmodel has received a lot of attention in biostatistics.The advantage of jointmodel is that it can jointly consider survival analysis data and longitudinal data on patient survival risk,because there is often a link between survival data and longitudinal data,and joint consideration of both types of data can reduce the residuals of model estimation and improve the efficiency of statistical inference.Much work has focused on the linear association structure between longitudinal and survival submodels,however,in practice both longitudinal biomarkers and patients’ instantaneous risk are related to time-varying covariates,and the linear association structure is difficult to handle the cases with time-varying covariates and censoring data.In this thesis,we propose a nonlinear association structured joint model based on random forest(RFJM),which can effectively solve the pain points of traditional linear structure by taking advantage of the characteristics of random forest that is good at handling complex data and missing data,and in addition,since RFJM is a joint structured model with nonlinearity,we developed a new parameter estimation method and dynamic prediction framework through the idea of optimization for RFJM based on the traditional joint model estimation method for dynamic prediction.Finally,to demonstrate the effectiveness of the proposed method,we performed empirical analysis on the PBC dataset and simulated dataset,and the final empirical results showed that(1)RFJM outperformed the traditional linear joint-structured joint model in terms of model prediction performance measured by the dynamic prediction metrics AUC and PE score;(2)RFJM can handle time-varying covariates better than traditional models,and the more complex the variables,the better the results;(3)RFJM also performs better than traditional models when facing data sets with more censored values.