Buckley-James Generalized Additive Model With Ultra-high Dimensional Data

In biomedical statistics,the study of censored ultra-high-dimensional data has always been difficult.Based on the sparsity assumption,it is indispensable to perform feature screening on ultra-high-dimensional data.Using a variety of feature screening methods to extend the prediction of survival time from a simple linear model to an additive model can further improve the accuracy and interpretability of the prediction.The data comes from a study related to diffuse large B-cell lymphoma(DLBCL)in the GEO database.It contains 414 patient samples,and their 54675 gene characteristics.In addition,the survival time T in the data is the right censored data,and the censoring ratio is 60%.In terms of feature screening,we consider the variation of explanatory variables and correlation between them,and use weighted p-value to choose variables.In the accelerating failure model,we take the Buckley-James estimation to fill in the censored survival time to obtain a robust estimation of the complete survival time T.The results show that the predictive accuracy of the additive accelerated failure time model is better compared with the traditional Cox model and the random survival forest.The calculation time of additive AFT model is significantly shorter than that of the random survival forest.