1) The document discusses injecting domain knowledge from ontologies into electronic medical records to improve prediction of patient hospitalization. It evaluates adding concepts from sources like ICPC2, DBpedia, Wikidata, and NDF-RT to bag-of-words representations used in machine learning.
2) The addition of domain knowledge, especially hierarchical ATC drug codes and focused DBpedia concepts, improved prediction performance according to the F-measure metric. Combining domain knowledge sources worked best.
3) Support vector machines with linear kernels, logistic regression, and random forests were tested on representations combining text and concepts. Random forests and logistic regression with domain knowledge provided the best hospitalization predictions.