Clinical Text Mining in Spanish Enhanced by Negation Detection and Named Entity Recognition

Abstract

Automatic identification of negation, uncertainty, and named entities are tasks of vital importance for clinical text mining. While several works have been published in English, only in recent years Spanish cases have been considered. In this work, we present a transfer learning framework based on a RoBERTa model pre-trained with biomedical documents and on multilingual BERT to identify diseases and organisms mentions as well as negations and uncertainty cues and scopes as a sequence labeling problem, utilizing the fact clinical datasets in Spanish for these four tasks. Our approach achieves results comparable to the state-of-the-art organism mentions identification and negation identification, competitive results in identifying diseases, and establishing state-of-the-art for uncertainty identification. Additionally, to remedy the lack of a unified dataset for the four tasks addressed, models to tackle them have been integrated into a web application that we built to allow effective clinical text mining in Spanish. The source code of this work is publicly available as well as the web application.