This document discusses using representation learning techniques in medical documents. It describes how word embeddings and document embeddings can be used to represent medical text in vector spaces to perform tasks like sentence classification. The document outlines an approach using convolutional neural networks (CNNs) to classify medical publications into 13 topics based on 27.4 million words from PubMed publications. CNNs were able to achieve good classification performance, but the authors note limitations in the dataset and opportunities for future work, such as using more complex models and developing patient representations from their records.