Feature selection for survival analysis

Details

Supervisors

Dupont, Pierre ; D'Ambrosio, Roberto

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

Survival analysis is at the core of clinical trials to make an accurate prognosis of patient status, and a base to decide appropriate treatments. Survival analysis is the class of statistical methods used to study the occurrence and timing of events, such as the development of metastasis in a cancer, relapse or a cancer related death. In this context, feature selection aims at finding a subset of relevant covariates among those provided by clinicians for specific disease like colon or liver cancer. These variables (the covariates) may be of any kind that clinicians saw fit, for instance the age, the sex, or some gene expression. The goal of feature selection is to identify a subset of those covariates that is informative of the disease progression. Several methods exist to perform this task, but none is known to outperform every other in both robustness and performances. In particular, some methods provide unstable feature selection, as the chosen covariates may change drastically when including a few more patients. For a clinician wanting to determine the treatment to prescribe to a patient, a preferred solution is to look at a few key factors, which is much simpler and faster, as well as much cheaper than having to measure gene expression levels of thousands genes. This can be achieved through feature selection, as it can substantially reduce the number of covariates deemed predictive of survival. What's more, selecting features that are more closely related to the disease under study will help clinicians understand the biological processes underlying it. The objectives of this master thesis are to design a new feature selection technique that is both robust and with high performances on survival data.