Identification of survival models that are prognostic across cohorts and stable regarding variable selection with methods of model-based optimization

点击翻译按钮获取中文摘要

The goal is the development of new statistical methods for the identification of survival models that are prognostic across cohorts and stable regarding variable selection. There is a large number of prediction methods for survival data based on clinical and high-dimensional genetic data, and the best model often depends on the patient cohort used. With modern methods of model-based optimization the best models for individual cohorts can be determined efficiently and with significantly reduced run times. The resulting models have two drawbacks. First, they are specialized for the learning cohort and thus often lack of high prediction accuracy on independent cohorts. Second, due to the inherent redundancy in the genetic measurements often different variables are selected on different cohorts despite the same biomedical question. For a better generalizability in this project methods are developed for the identification of models that are at the same time prognostic across cohorts and stable regarding variable selection. A model identified in such a multi-objective approach must then be compared with the cohort-specialized models in order to evaluate the loss in prediction accuracy due to the additional stability criteria. The reproducibility can be significantly improved by integrating public open access experiment databases, since comparison studies of prediction methods then can be extended collaboratively and transparently. The result of this work will be models that are prognostically relevant with a stable variable selection that allows a biological interpretation of the genetic features.