Model-free Conditional Feature Screening For Case ? Interval-censored Failure Time Data With Ultrahigh Dimensional Covariates

Along with the rise of technologies such as artificial intelligence and precision medicine,ultrahigh dimensional covariates with interval-censored data as the survival outcome are widely appearing in clinical trials,medical images and other research fields.Most of the previous proposed ultrahigh dimensional screening methods are for cases where the response variables can be exactly observed or right censoring exists,but the complexity of the interval-censored data forms makes it difficult for the existing methods to be directly applied.The traditional methods based on sure independent screening have some limitations due to ignoring the case where strong correlations among covariates.In order to effectively reduce the dimensionality of the ultrahigh dimensional feature space while avoiding the problems that traditional methods are prone to,we provide a feature screening approach based on conditional screening for ultrahigh dimensional case ? interval-censored data by introducing known information into the conditional survival function of interval-censored data,which could measure the correlation between each covariate and the distribution of response variable under a given condition.To avoid specific model assumptions,the sample estimators of proposed method are obtained by using Nadaraya-Watson kernel estimation and weighted Turnbull's selfconsistency algorithm,and we theoretically demonstrate that the screening procedure satisfies the sure screening property and ranking consistency property.In other words,it can effectively distinguish important and non-significant features after a pre-given model size.The numerical simulation part evaluates the performance of proposed screening statistics based on different models and setting conditions under a limited sample,also considers two special cases when the known information is not an important variable and when no threshold is given in advance.Numerical results show that our approach can select all significant features with high probability when there is a strong correlation between the conditional and unimportant variables.Besides,the performance is significantly improved with the increase of the sample size.The method is also applied to the adjusted DLBCL real dataset and compared with existing studies to illustrate its feasibility.Compared with existing methodologies,our method is less susceptible to the limitations of model assumptions and bandwidth,which could effectively overcome the loss of traditional marginal screening ideas.