Prediction of time to event for censored data: ridge regression with linear constraints in kernel space

We propose a new method for analyzing time to event in case of partially censored data and compare its performance for the particular task of breast cancer metastasis prediction with the performance of several known methods trained on the same data. In our approach, we use ridge regression for uncensored data, treating censored samples as constraints. Instead of initial feature space we use feature space defined by a kernel function. We reduce dimensionality by using coefficient of variation for each regression coefficient as a criterion for eliminating corresponding dimension.