University of Toronto, Canada
Title: Recursive partitioning method on survival outcomes for personalized medicine
Wei Xu is a Principal Biostatistician at the Princess Margaret Cancer Center, a Scientist at the Ontario Cancer Institute, and a faculty at Dalla Lana School of Public Health, University of Toronto. Dr. Xu’s research interests focus on clinical trial design and methodology, statistical genetics, and cancer translational research. He has been involved in various clinical studies and human genetic studies. As Principal-Investigator or co-Investigator on multiple research studies, he has been involved in study design, data administration, statistical modeling and analysis on different clinical research and human genetic studies. So far, he has been published over 150 peer-reviewed papers in high impact journals of statistics, bioinformatics, medical science, and human genetics.
Abstract Background: A general method to create adjusted recursive partitioning (tree-based) model on survival outcomes is developed. Prognostic survival trees have historically been used to automatically uncover complicated GxG and GxE interactions. However scientists often want to uncover this structure while adjusting for confounding factors that are not of direct interest. Interaction survival trees can automatically identify the best treatment choice for patients and are a promising model to enable personalized medicine, but simulations to assess their performance on the high dimensional data found in personalized medicine have not been conducted. Methods: We develop a general framework to adjust for confounding factors in prognostic and interaction survival trees. These factors are numerous in practice and can include age, gender, study site in a randomized multicenter clinical trial, and the principal components of ancestry difference to control for population stratification in genetic studies. Results: Extensive simulations show the performance of our methods under various true tree structures. Our methods are shown to be well controlled under the null with only a 1.4-8.4% chance to build a spurious tree when none should be made. Under the alternative, the power to build the correct tree is robust to the large dimensional covariate space found in personalized medicine, dropping less than 2% when going from 10 to 1,000 potential splits. We applied our adjusted interaction tree on a randomized clinical trial study on head and neck cancer patients. The novel method successfully identify subgroups of head and neck cancer patients that respond positively to having antioxidant vitamins added to their treatment regime. These subgroups are based on the patients genetic signature and are adjusted for population stratification. Conclusions: We have demonstrated that our adjusted survival tree method can create prognostic and interaction survival trees that are adjusted for confounders not of direct interest. We have also shown that our adjusted interaction survival trees
Speaker PPTsDownload PPT