Hybrid medical named entity recognition using document structure and surrounding context

Nowadays, there is a huge amount of electronic medical documents created in natural language by medical specialists, containing useful information needed for several medical tasks. However, reading these documents to get some specific information is a too tiring task. Thus, extracting information automatically became an essential and a challenging task, especially Named Entity Recognition (NER). NER is crucial for extracting valuable information used in various medical tasks such as clinical decision support and drug safety surveillance. Capturing sufficient context is important for an efficient NER. In the literature, some important context information are not well exploited. Usually, a standard sequence segmentation is used, such as sentence segmentation, which may can’t cover sufficient context. In this paper, we propose a supervised NER method, called MedSINE (Medical Section Identification to enhance the Named Entity tagging), which is based on sequence tagging task using Bidirectional Long Short-Term Memory neural network with Conditional Random Field (BiLSTM-CRF). For that, we exploit layout information to segment the text on chunk sequences and to extract the parent sections of each word as features to provide sufficient context. In addition, we have used a clinical Bidirectional Encoder Representations from Transformers (BERT) word embedding, Part of Speech (PoS), and entity surrounding sequence features. Experiments were conducted on a manually annotated dataset of real Summary of Product Characteristics (SmPC) medical documents in PDF format and on the Colorado Richly Annotated Full Text (CRAFT) corpus. Our model achieved an F1-measure of 89.49%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$89.49\%$$\end{document} and 73.52%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$73.52\%$$\end{document} in terms of strict matching evaluation using the SmPC and CRAFT datasets, respectively. The results show that employing the sequence of parent sections improves the F1-measure by 4.71%\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$4.71\%$$\end{document} in terms of strict matching evaluation.