Overall survival prediction of glioma patients using genomics

Abstract

Overall survival prediction is a vital task that will lead for better patient management in clinical practise. Existing approaches mainly focus on imaging based survival prediction, which is non invasive, and thus, easier to be implemented at the initial diagnosis stages. However, the advancements in the DNA/RNA technologies has given access to genomic and transcriptomic profiles of the gliomas, that directly reflect the molecular level alterations. Thus, in this work we mainly focus on using transcriptomic profiles for survival prediction, an area that has not been widely analysed yet for survival prediction. We utilize the gene expression and mutation profiles, while augmenting the recent Artificial Intelligence approaches, such as deep probabilistic programming and multi task learning for prognosis prediction. Thereby we do not just focus on the application, we also contribute with novel learning paradigms to improve the classification task performances. Nonetheless, we also focus on proposing a novel loss function, since architectural wise the state of art performance has been achieved for classification tasks. In addition, we also investigate ability to employ radiomics, for subtype classification, that is also associated with survival. Since subtypes mainly rely on the genomic alterations, we found it useful to focus on imaging features ability to predict prognosis of glioma.