Medical entity recognition and negation extraction: Assessment of NegEx on health records in Spanish

Abstract

This work focuses on biomedical text mining. The core of this work is to make a step ahead in the negation detection of biomedical entities on Electronic Health Records (EHRs), where the detection of non-negated entities is as important as the identification of negated entities. For instance, the identification of a negated entity as factual, can produce diagnostic errors in decision support systems. Negated entity recognition tackles two tasks: (1) entity recognition; (2) entity classification as negated or not. To identify negations, in the literature rule-based and machine-learning techniques have been used. This paper presents an adaptation of the rule-based system NegEx, which uses exact-matching for the aforementioned tasks. Our contribution consist in assessing the aforementioned two tasks and explored alternatives for each of them, in such a way that the negation detection improves when the entity recognition is able to detect more entities correctly. The evaluation was carried out within a real domain of 75 EHRs written in Spanish obtaining an f-measure of 76.2 for entity recognition and 73.8 for negation detection.