Role of preferred terminology in the classification of medical reports

We explore the role of the unified medical language system metathesaurus preferred terms in multi-label classification of medical reports. We use preferred terms to represent lexically-disparate but semantically-equivalent medical concepts. We experiment with two classifiers: one based upon Lucene and the other based upon BoosTexter. We find the performance of these systems fails to show consistent improvement when their input is enriched with preferred terminology. Our analysis shows that in our data, some reports can be correctly classified not only in the absence of preferred terms, but in some cases, in the absence of direct references to the diseases themselves.