Abbreviation Identification in Clinical Notes with Level-wise Feature Engineering and Supervised Learning

Abstract

Nowadays, electronic medical records get more popular and significant in medical, biomedical, and healthcare research activities. Their popularity and significance lead to a growing need for sharing and utilizing them from the outside. However, explicit noises in the shared records might hinder users in their efforts to understand and consume the records. One kind of explicit noises that has a strong impact on the readability of the records is a set of abbreviations written in free text in the records because of writing-time saving and record simplification. Therefore, automatically identifying abbreviations and replacing them with their correct long forms are necessary for enhancing their readability and further their sharability. In this paper, our work concentrates on abbreviation identification to lay the foundations for de-noising clinical text with abbreviation resolution. Our proposed solution to abbreviation identification is general, practical, simple but effective with level-wise feature engineering and a supervised learning mechanism. We do level-wise feature engineering to characterize each token that is either an abbreviation or a non-abbreviation at the token, sentence, and note levels to formulate a comprehensive vector representation in a vector space. After that, many open options can be made to build an abbreviation identifier in a supervised learning mechanism and the resulting identifier can be used for automatic abbreviation identification in clinical text of the electronic medical records. Experimental results on various real clinical note types have confirmed the effectiveness of our solution with high accuracy, precision, recall, and F-measure for abbreviation identification.