ULisboa: Identification and Classification of Medical Concepts

Abstract

This paper describes our participation on Task 7 of SemEval 2014, which focused on the recognition and disambiguation of medical concepts. We used an adapted version of the Stanford NER system to train CRF models to recognize textual spans denoting diseases and disorders, within clinical notes. We considered an encoding that accounts with noncontinuous entities, together with a rich set of features (i) based on domain specific lexicons like SNOMED CT, or (ii) leveraging Brown clusters inferred from a large collection of clinical texts. Together with this recognition mechanism, we used a heuristic similarity search method, to assign an unambiguous identifier to each concept recognized in the text.Our best run on Task A (i.e., in the recognition of medical concepts in the text) achieved an F-measure of 0.705 in the strict evaluation mode, and a promising F-measure of 0.862 in the relaxed mode, with a precision of 0.914. For Task B (i.e., the disambiguation of the recognized concepts), we achieved less promising results, with an accuracy of 0.405 in the strict mode, and of 0.615 in the relaxed mode.

ULisboa: Recognition and Normalization of Medical Concepts

“…We applied the same type of approach used in our system from last year (Leal et al, 2014) for entity recognition. The Stanford NER software (Finkel et al, 2005) was used to train Conditional Random Fields (CRF) models using labelled data as input.…”

“…All input text had to be tokenized and encoded according to a named entity recognition scheme that encodes entities as token classifications. To be able to recognize non-continuous entities, an SBIEON (Leal et al, 2014) encoding was used. Besides the tags defined in the SBIEO encoding , a new tag N was added to identify words that do not belong to the entity but are inside the continuous span that contains the recognized entity.…”

“…The entity recognition module is based on the Stanford NER tool (Finkel et al, 2005), and it uses CRF models trained on annotated biomedical notes. The module tags the text according to an SBIEON encoding of entities as token classes, supporting the recognition of non-continuous entities (Leal et al, 2014). We relied on features based on Brown clusters and domain specific lexicons.…”

SemEval-2014 Task 7: Analysis of Clinical Text

“…A novel aspect of the SemEval-2014 shared task that differentiates it from the ShARE/CLEF task-other than the fact that it used more data and a new test set-is the fact that SemEval-2014 allowed the use of a much larger set of unlabeled MIMIC notes to inform the models. Surprisingly, only two of the systems (ULisboa (Leal et al, 2014) and UniPi (Attardi et al, 2014)) used the unlabeled MIMIC corpus to generalize the lexical features. Another team-UTH CCB (Zhang et al, 2014)-used off-the-shelf Brown clusters 10 as opposed to training them on the unlabeled MIMIC II data.…”

“…The prefix H indicates that the word or word sequence is the shared head, and the prefix D indicates otherwise. Another intermediate approach used by the ULisboa team (Leal et al, 2014) with the tagset-S, B, I, O, E and N. Here, S represents the single token entity to be recognized, E represents the end of an entity (which is part of one of the prior IOB variations) and an N tag to identify non-contiguous mentions. They don't provide an explicit example usage of this tag set in their paper.…”