| In the field of survival analysis,there is a common phenomenon of censored data,where only one observation is made on the research object and only the exact occurrence time of the event of interest is known before or after the observation time,which is called interval type I censored data,also known as current status data.In the medical field,medical images become an important part of medical data with the characteristics of information visualization and auxiliary diagnosis of clinical results.In order to examine the relationship between image data and survival time,scalar-on-image regression models can usually be established.However,the dimensions of general medical image data are high and there is spatial correlation,with only some blocks related to survival time.So for high-dimensional data with complex spatial correlation structures,improving the interpretability of the model through variable selection has become a research hotspot.Therefore,this article conducts spatial Bayesian variable selection for Accelerated Failure Time model and Proportional Hazards model based on current status data,and specifically conducts the following two parts of research:The first part discusses the selection of spatial Bayesian variables for high-dimensional scalar-on-image Accelerated Failure Time models under current status data.First,under the current status data,the image covariates are combined into the Accelerated Failure Time model,and the Ising-DP spike-and-slab prior is established by introducing Ising prior and Dirichlet Process prior.Clustering grouping and variable selection are carried out for the image covariates with the same effect and greater impact on the response variables.In order to simplify the posterior calculation,the Markov Chain Monte Carlo algorithm combined with data enhancement is used for sampling estimation.Then,through simulation research,verify the effectiveness and feasibility of the Ising-DP spike-and-slab method.Finally,the model method is applied to the classic Alzheimer’s Disease Neuroimaging Initiative dataset,and voxel regions that have a greater impact on the onset time of Alzheimer’s Disease in Hippocampus are selected.The second part discusses the selection of spatial Bayesian variables for highdimensional scalar-on-image Proportional Hazards model under current status data.Firstly,a Proportional Hazards model is established for survival time and image covariates.Monotonic splines are used to process the unspecified cumulative hazards function.Under the Bayesian framework,a mixed prior of Ising-DP spike-and-slab regression coefficients is given to select spatial variables for image covariates.In order to reduce the complexity of the current status data and simplify the posterior distribution calculation,using the relationship between Proportional Hazards model and non-homogeneous Poisson Process and the additivity of poisson random variables,the likelihood function is expanded with twostage poisson data,and the ARMS algorithm and Gibbs algorithm are used to obtain parameter estimation.Then,based on multiple simulation studies,it was found that compared to the independent Bernoulli Dirichlet Process method,the Ising-DP spike-andslab method not only has more accurate selection,but also has higher efficiency in highdimensional environments.Finally,the model method is applied to the pretreated dataset of Alzheimer’s Disease.The results show that the Ising-DP spike-and-slab method is also feasible in the semi parametric Proportional Hazards model,and verify the objective fact that the atrophy characteristics of the Hippocampus have a greater impact on the onset time of Alzheimer’s Disease. |
