Constructing survival curves from censored data with machine learning methods

Aljaž Košmerlj (2008) Constructing survival curves from censored data with machine learning methods. EngD thesis.

PDF Download (466Kb)

Abstract

In the present thesis I introduce and evaluate a new machine learning method for estimating survival functions from survival analysis data. Firstly, I describe the field of survival analysis and the problems it deals with. I introduce and define the basic terms of survival analysis, like survival function and survival curve. I also define censored data, a speciallity of survival analysis data, and explain their importance and the learning problems they cause. As a reference method I describe the Kaplan-Meier estimator, a well-known statistical method for estimating survival curves, that serves as a conceptual basis for the new proposed method. I close the introduction with a short overview of the advances of machine learning in the field of survival analysis, concluding that so far there are no well established meachine learning methods in this field. I continue with an in depth description of the proposed method and it's potential advantages. To test the new method thoroughly I start with a series of tests on artificially generated data from a physics domain. The new method proves itself useful and can match the accuracy of the Kaplan-Meier estimator. I discuss the problem of nonmonotonic survival curve estimations, that can be obtained using the proposed method. All the tests are repeated on a set of real medical data describing the prognostic value of protein markers for survival of metastatic breast cancer patients. The results further confirm the proposed method as useful. In conclusion I present the possibilities of improving the proposed method and suggest other prospects of using machine learning techniques in survival analysis.

Actions (login required)