In this study, we investigate different strategies for assigning MeSH (Medical Subject Headings) terms to clinical guidelines using machine learning. Features based on words in titles and abstracts are investigated and compared to features based on topics assigned to references cited by the guidelines. Two of the feature engineering strategies utilize word embeddings produced by recent models based on in the distributional hypothesis, called word2vecand fastText. The evaluation results show that reference-based strategies tend to yield a higher recall and F1 scores for MeSH terms with a sufficient amount of training instances, whereas title and abstract based features yield a higher precision.