Classifying Short Clinical Notes: An Unsupervised Approach

Abstract

A mandatory task in Sweden is the reporting of clinical procedures with a specially assigned code based on the procedure. It is both time-consuming and troublesome for medical personnel since more than 10,000 codes exist. By automating this task, it is possible to both save time of the personnel and money within the healthcare industry. This master thesis explores an alternative way of classifying short clinical notes through unsupervised methods when quality labelled data is not available. By combining advances within NLP, utilising word embeddings and incorporating additional knowledge into the data, a classifier which do not rely on labelled data is presented. Instead of learning by examples as supervised methods, the classifier manages to find semantic similarities between clinical notes and the description of the different codes, making it intuitively similar to how we humans would classify a code.