|
Abstract:
|
A well-defined staging system constructed by orderly grouping the cancer characteristics plays a crucial role in cancer diagnosis, treatment and prognosis. It is common to have clinical trials collecting the same covariates and outcomes, but conducted on different population. Data from these trials usually suffers from limited sample size or even incomplete feature table when they are separate. But it is also difficult to combine them, because of heterogeneity. We propose to use a group fused Lasso penalty to complete the two tasks at the same time: orderly grouping the cancer features is achieved by the stage-enforcing property of the fused Lasso, and borrowing strength across data sets is fulfilled through properly selecting the "groups" of the coefficients from different data sets. We demonstrate the advantageous performance of our method by a simulation study. We also apply it to a breast cancer data with survival outcome of individual level.
|
