Improving survival prediction using a novel feature selection and feature reduction framework based on the integration of clinical and molecular data.

The accurate prediction of a cancer patient's risk of progression or death can guide clinicians in the selection of treatment and help patients in planning personal affairs. Predictive models based on patient-level data represent a tool for determining risk. Ideally, predictive models will use multiple sources of data (e.g., clinical, demographic, molecular, etc.). However, there are many challenges associated with data integration, such as overfitting and redundant features. In this paper we aim to address those challenges through the development of a novel feature selection and feature reduction framework that can handle correlated data. Our method begins by computing a survival distance score for gene expression, which in combination with a score for clinical independence, results in the selection of highly predictive genes that are non-redundant with clinical features. The survival distance score is a measure of variation of gene expression over time, weighted by the variance of the gene expression over all patients. Selected genes, in combination with clinical data, are used to build a predictive model for survival. We benchmark our approach against commonly used methods, namely lasso- as well as ridge-penalized Cox proportional hazards models, using three publicly available cancer data sets: kidney cancer (521 samples), lung cancer (454 samples) and bladder cancer (335 samples). Across all data sets, our approach built on the training set outperformed the clinical data alone in the test set in terms of predictive power with a c.Index of 0.773 vs 0.755 for kidney cancer, 0.695 vs 0.664 for lung cancer and 0.648 vs 0.636 for bladder cancer. Further, we were able to show increased predictive performance of our method compared to lasso-penalized models fit to both gene expression and clinical data, which had a c.Index of 0.767, 0.677, and 0.645, as well as increased or comparable predictive power compared to ridge models, which had a c.Index of 0.773, 0.668 and 0.650 for the kidney, lung, and bladder cancer data sets, respectively. Therefore, our score for clinical independence improves prognostic performance as compared to modeling approaches that do not consider combining non-redundant data. Future work will concentrate on optimizing the survival distance score in order to achieve improved results for all types of cancer.