This document describes research on automatically classifying the polarity (positive/negative recommendations) of sentences in evidence-based medicine (EBM) documents. The researchers annotated sentences from clinical questions and trained a support vector machine classifier to determine polarity. Initial results showed over 80% accuracy is achievable, and using context features like the intervention being evaluated improved performance over prior work. More training data is needed to generate bottom-line recommendations, but preliminary results were promising. The goal is to help summarize EBM documents by determining which interventions are recommended or not recommended.