Testing Contextualized Word Embeddings to Improve NER in Spanish Clinical Case Narratives

Background: In the Big Data era there is an increasing need to fully exploit and analyse the huge quantity of information available about health. Natural Language Processing (NLP) technologies can contribute to extract relevant information from unstructured data contained in Electronic Health Records (EHR) such as clinical notes, patient’s discharge summaries and radiology reports among others. Extracted information could help in health-related decision making processes. Named entity recognition (NER) devoted to detect important concepts in texts (diseases, symptoms, drugs, etc.) is a crucial task in information extraction processes especially in languages other than English. In this work, we develop a deep learning-based NLP pipeline for biomedical entity extraction in Spanish clinical narrative. Methods: We explore the use of contextualized word embeddings to enhance named entity recognition in Spanish language clinical text, particularly of pharmacological substances, compounds, and proteins. Various combinations of word and sense embeddings were tested on the evaluation corpus of the PharmacoNER 2019 task, the Spanish Clinical Case Corpus (SPACCC). This data set consists of clinical case sections derived from open access Spanish-language medical publications. Results: NER system integrates in-domain pre-trained Flair and FastText word embeddings, byte-pairwise encoded and the bi-LSTM-based character word embeddings. The system yielded the best performance measure with F-score of 90.84%. Error analysis showed that the main source of errors for the best model is the newly detected false positive entities with the half of that amount of errors belonged to longer than the actual ones detected entities. Conclusions: Our study shows that our deep-learning-based system with domain-specific contextualized embeddings coupled with stacking of complementary embeddings yields superior performance over the system with integrated standard and general-domain word embeddings. With this system, we achieve performance competitive with the state-of-the-art.