A Study of Patient-Specific Prognosis of Ovarian Cancer

Abstract

Current cancer prognosis are based on broad population averages statements. This thesis, focused on ovarian cancer, aims to estimate patients survival time. Different Neural Networks are tested on a medical dataset containing physiological information on patients. First predictions on the survival time are obtained by standard point estimators such as Multilayer Perceptrons (MLP) and Radial Basis Function (RBF) networks. But as the results are quite disappointing, a novel estimation technique is introduced: Mixture Density Networks (MDN). The MDN method provides a probabilistic model for the estimation which cannot be obtained by others methods. Hence we obtained the full distribution of the probabilities of the survival time and discovered that it is highly multimodal, so no reliable prediction can be made. Indeed, the error rate obtained with the best model is about 70 %. Finally, some attempts at classifying patients into different classes of survival time are made, and the results are quite surprising as the Neural Networks can only distinguish censored patient and patients with deadly outcome.