Automated Annotation for Medical Discharge Summaries: A Preliminary Study

Presumably, with sufficient training data, it would be possible to build a classifier for each of the twenty predetermined complication labels. However since our training set consists of only a hundred and forty patients in total, and 14 of the 20 labels were associated with 10 or fewer patients, this approach was clearly infeasible. Similarly for the second task, there were only twenty patients labeled with additional complications. Instead of relying solely on manually annotated training data, we leverage it in order to devolop a pattern matcher that automatically generates more “labeled” data. This output is then used to train a classifier that recognizes whether or not a particular sentence contains some mention of a complication. The classifier is feature engineered to capture the general characteristics (linguistic and otherwise) of any complication mention, that is, it captures instances from a much broader set than the twenty complications the pattern matcher was developed for. 2 Background