Deep learning has achieved remarkable results in the areas of computer vision, speech recognition, natural language processing and most recently, even playing Go. The application of deep-learning to problems in healthcare, however, has gained attention only in recent years, and it's ultimate place at the bedside remains a topic of skeptical discussion. While there is a growing academic interest in the application of Machine Learning (ML) techniques to clinical problems, many in the clinical community see little incentive to upgrade from simpler methods, such as logistic regression, to deep learning. Logistic regression, after all, provides odds ratios, p-values and confidence intervals that allow for ease of interpretation, while deep nets are often seen as `black-boxes' which are difficult to understand and, as of yet, have not demonstrated performance levels far exceeding their simpler counterparts. If deep learning is to ever take a place at the bedside, it will require studies which (1) showcase the performance of deep-learning methods relative to other approaches and (2) interpret the relationships between network structure, model performance, features and outcomes. We have chosen these two requirements as the goal of this study. In our investigation, we utilized a publicly available EMR dataset of over 32,000 intensive care unit patients and trained a Deep Belief Network (DBN) to predict patient mortality at discharge. Utilizing an evolutionary algorithm, we demonstrate automated topology selection for DBNs. We demonstrate that with the correct topology selection, DBNs can achieve better prediction performance compared to several bench-marking methods.