Support Vector Machines for Censored Data

University Of North Carolina At Chapel Hill, Chapel Hill NC

Investigators

Abstract

Recent advances in medical research, including those related to the study of the human genome, have led to the development of personalized medicine. Personalized medicine describes medical treatment that is tailored to a patient based on the patient's genetic profile and other personal biomedical information. Developing personalized medical treatment regimens is challenging since it involves learning from high-dimensional patient data. Moreover, data from personalized-medicine clinical studies are typically subject to censoring, i.e., the data are not fully observed, due, for example, to patients dropping out during the course of a study. While statistical analysis of censored data is a well-developed area, most of the existing statistical tools were developed under restrictive assumptions that often do not hold for high-dimensional data settings. It is therefore important to develop an approach for analysis of censored data that can be applied to today's high-dimensional data sets. In this project we will develop novel machine learning techniques that can handle both high-dimensional and censored data. The algorithms that we will develop will be applicable not only in the field of personalized medicine, but also in other disciplines in which high-dimensional censored data are common, such as engineering, economics, and sociology. In this research, we will extend the framework of support vector machines (SVMs) to censored data. First, we will develop support vector learning techniques for different types of censored data, including right censored data, interval censoring, current status data, and multistage decision problems with censored data. We will then study the theoretical properties of these estimators, including their finite-sample properties and their asymptotic behavior. For this goal we will develop novel methodology, including new finite-sample tools for censored data. Finally, we will apply the tools that we develop to real-world data. We will compare the proposed learning methods with existing methods using theoretical tools, simulation, and analysis of real-world data. Finally, we will develop software for each of the different algorithms that we study. This software will be developed such that it can be integrated into existing machine learning software.