An Experiment in Automatic Classification of Pathological Reports

Abstract

Abstract. Medical reports are predominantly written in natural language; as such they are not computer-accessible. A common way to make medical narrative accessible to automated systems is by assigning 'computer-understandable' keywords from a controlled vocabulary. Experts usually perform this task by hand. In this paper, we investigate methods to support or automate this type of medical classification. We report on experiments using the PALGA data set, a collection of 14 million pathological reports, each of which has been classified by a domain expert. We describe methods for automatically categorizing the documents in this data set in an accurate way. In order to evaluate the proposed automatic classification approaches, we compare their output with that of two additional human annotators. While the automatic system performs well in comparison with humans, the inconsistencies within the annotated data constrain the maximum attainable performance.