Systematic investigation of hyperparameters on performance of deep neural networks: Application to ovarian cancer phenotypes

Abstract

The application of Deep Neural Networks (DNNs) to medicine has recently emerged as a major approach for prognosis. Many medical researchers have expected that the use of the DNN algorithms would provide higher prediction results for their analysis. However, while these applications are currently underway for medical imaging data, they are not yet optimised for clinicopathologic data, with two-dimensional input space. One such challenge is the difficulty of applying deep learning to optimise hyperparameters, i.e., learning DNN models for more accurate prediction results. In this study, we identified parameters having a greater impact on predictive power, by applying DNNs to clinicopathologic data. Specifically, we predicted therapeutic response to platinum-based chemotherapy, based on data from 710 epithelial ovarian cancer patients. Predictive performance was measured by Area Under the Curve (AUC) after optimising six hyperparameters, including the number of hidden layers, number of hidden units, optimisation algorithm, weight initialisation, activation function for hidden layers, and dropout rate. By identifying the significant main effects, and interaction effects, of these hyperparameters on clinical prediction, we successfully determined combinations of hyperparameters contributing to higher predictive power. These approaches have ramifications for assessing therapeutic response to numerous treatments for various pathologies.