MeSHx-Notes: Web-System for Clinical Notes Information Extraction

Abstract

We present MeSHx-Notes, MeSH eXtended for clinical notes, a multi-language web system based on the Django framework to present information selected in clinical notes. MeSHx-Notes extends Medical Subject Headings (MeSH) terms with Word Embeddings with similar semantic/syntactic words. Since MeSH is available for 15 languages, MeSHx-Notes is easily extendable by replacing the MeSH thesaurus with the target language. In this demo, we show examples with Portuguese and English.

Keywords

• Multi-language
• Web System
• Clinical Notes
• Information Extraction
• Word Embeddings
• MeSH

Introduction

Electronic Health Records (EHR) play an important role in hospital environments, bringing many bene ts in terms of patient safety, e ectiveness and e ciency of care, and patient satisfaction [ 1 ]. Records of health care practices in hospitals generate a rich and large amount of patient information and an intrinsic relation between symptoms, diseases, drug interaction, and diagnoses that may be used for many purposes [ 4 ].

This study aims to help healthcare professionals concerning the understanding of what has been informed by a clinical note. This is possible through the use of Natural Language Processing (NLP), combined with the MeSH dictionary. The system consists of a web application that exhibits the meaning and the related words of the main terms used in clinical notes, thus enhancing the understanding of what is reported.

Other systems, such as cTAKES [ 3 ], rely on several UMLS sources for English to provide several information from clinical notes. We focus on developing a userfriendly and easy-handling UI through a web application, portable for languages other than English, using a language-speci c MeSH thesaurus.

In this context, we present an easy-to-use system that provides users with extra knowledge of the information given in clinical notes, which can be used by anyone with access to the internet. The system consists of a web application that receives clinical notes, identi es the main terms, and then, returns their de nition, similar words and a link to the MeSH dictionary. Its development is based on Python, Django, Pandas, Bootstrap, JQuery, Word Embeddings, XPath, and the MeSH thesaurus. 2.1

Data Source

Three resources are used to develop MeSHx-Notes, as described below. In addition, we describe the process to generate the word embedding vectors. Electronic Health Records The Portuguese dataset was obtained from Hospital Nossa Senhora da Conceic~ao (HNSC). The English dataset was obtained from i2b2 Challenge [ 5 ] from 2008 to 2012. It is a set of nine datasets from several shared tasks promoted by Informatics for Integrating Biology and the Bedside (i2b2).

Medical Subject Headings (MeSH) MeSH is a "controlled vocabulary" Metathesaurus, developed by the National Library of Medicine (NLM). As of 2013, MeSH has 54,935 entries where each entry has a unique tree number and consists of 26,851 main headings and 213,000 entry terms that increase the power of classi cation of medical documents.

Word Embeddings Word vectors are a way of mapping words in a numerical space. A latent syntactic/semantic vector for each word is induced from a large unlabeled corpus. The Portuguese and English model for the word embeddings was trained with Word2Vec [ 2 ]. For the Portuguese version, we used 21 million sentences from HNSC's medical records, trained with 50 dimensions per word and 100 minimum word count. This training resulted in 63 thousand word vectors used as a semantic model in the neural network below. For the English version, we used 171 thousand sentences from the i2b2 challenge dataset, trained with 50 dimensions and 10 minimum word count, resulting in 17 thousand word vectors. 2.2

Back-end

First, the MeSH dictionary is generated, using previously saved data in an XML le, containing ID, name, scope, terms, and quali er. The dictionary is enriched by identifying similar words using Word Embeddings, so that we provide a greater range of terms, which are stored in the terms eld. We consider higher similarity degrees to identify those words. An initial evaluation made by the authors resulted in 67% accuracy.

MeSHx-Notes: Web-System for Clinical Notes Information Extraction Heading Original Terms New Similar Terms Abdomem abdomem, belly abd, abdome...

Celecoxib celecoxib, celebrex norvasc, losartan...

Abscess abcesso, absceso abscess, abscesses...

In Table 1 we show some concepts that are commonly used in clinical notes. Each concept has a heading in the MeSh dictionary, its terms, and the new identi ed terms. For example, the heading "Abscesso" had "abcesso" and "absceso" as the original terms, and "abscess" and "abscesses" were added as new terms.

After that step, we read the clinical notes, using Pandas, in the web application, using Django as the development framework. Each word found in the dictionary is captured and the lists of original and new similar words are stored. 2.3

Front-end

When a clinical note is shown to the user, the words from the (enriched) dictionary are highlighted. These words are shown in di erent colors, according to the classes: medication, diagnosis, procedure or anatomy. We provide users with not only that, but also a navigation bar to go through all the desired clinical notes. This page is developed using JQuery and Bootstrap. MeSHx-Notes is presented for Portuguese and English samples. We use Word Embeddings for the dictionary expansion, in Portuguese and English.

In the web page, buttons are provided to navigate between clinical notes and to change the language. Besides, the clinical note description is given with data about the patient record and its modi cation date with a concomitant section of legends that are related to the classi cation of the terms. Nonetheless, identi ed words are underlined according to their classi cation, so that, when clicked, they show their technical name, ID, description, terms with similar meanings, and a link to the MeSH description website. 4

Conclusion and Further Work

MeSHx-Notes is able to provide, both for health professional and for non-specialists, a simple tool that enables a better understanding of the terms used in clinical notes in a clear, concise, accessible way. The source code is available on the project's Github page1, and the demo is found on the group's website2. As further work, we plan to use bigram and trigram embeddings to nd similar multi-word expressions.

Acknowledgments This work was partially supported by CAPES (Coordenac~ao de Aperfeicoamento de Pessoal de N vel Superior) Foundation (Brazil), PUCRS (Ponti cal Catholic University of Rio Grande do Sul) and UFRGS (Federal University of Rio Grande do Sul). 1 https://github.com/nlp-pucrs/meshx-notes 2 http://grupopln.inf.pucrs.br/meshx

References

1. Buntin , M.B. , Burke , M.F. , Hoaglin , M.C. , Blumenthal , D. : The bene ts of health information technology: a review of the recent literature shows predominantly positive results . Health a airs 30 ( 3 ), 464 { 471 ( 2011 )

2. Mikolov , T. , Sutskever , I. , Chen , K. , Corrado , G.S. , Dean , J. : Distributed representations of words and phrases and their compositionality . In: Advances in neural information processing systems . pp. 3111 { 3119 ( 2013 )

3. Savova , G.K. , Masanz , J.J. , Ogren , P.V. , Zheng , J. , Sohn , S. , Kipper-Schuler , K.C. , Chute , C.G. : Mayo clinical text analysis and knowledge extraction system (ctakes): architecture, component evaluation and applications . Journal of the American Medical Informatics Association 17 ( 5 ), 507 { 513 ( 2010 )

4. Silveira , M. , Nogueira , V.B. , Rodrigues , I. : Ferramentas e tecnologias para a integraca~o e extraca~o de informaca~o hospitalar . INF - Artigos em Livros de Actas/Proceedings ( 2015 )

5. Uzuner , O ., South , B.R. , Shen , S. , DuVall, S.L.: 2010 i2b2/ va challenge on concepts, assertions, and relations in clinical text . Journal of the American Medical Informatics Association 18 ( 5 ), 552 { 556 ( 2011 )