Predictive Modeling of Therapy Decisions in Metastatic Breast Cancer with Recurrent Neural Network Encoder and Multinomial Hierarchical Regression Decoder

Abstract

The increasing availability of novel health-related data sources —e.g., from molecular analysis, health Apps and electronic health records— might eventually overwhelm the physician, and the community is investigating analytics approaches that might be useful to support clinical decisions. In particular, the success of the latest developments in Deep Learning has demonstrated that machine learning models are capable of handling —and actually profiting from— high dimensional and possibly sequential data. In this work, we propose an encoder-decoder network approach to model the physician's therapy decisions. Our approach also provides physicians with a list of similar historical patient cases to support the recommended decisions. By using a combination of a Recurrent Neural Network Encoder and a Multinomial Hierarchical Regression Decoder, we specifically tackle two common challenges in modeling clinical data:First, the issue of handling episodic data of variable lengths and, second, the need to represent hierarchical decision procedures. We conduct experiments on a large real-world dataset collected from thousands of metastatic breast cancer patients and show that our model outperforms more traditional approaches.