Grant: Master's Award: Canada Graduate Scholarships
University of Toronto
$17,500.00
Improving word sense disambiguation in clinical narratives using deep learning
FY 2018-19
Summary
My proposal is focused on using deep learning to disambiguate abbreviations in clinical notes so that they are more comprehensive. Abbreviation disambiguation (AD) is challenging since abbreviations can have many senses. For example, ra can mean right atrium, rheumatoid arthritis or room air depending on its context. Correctly expanding medical terms is important to build accurate patient profiles, link notes to ontological concepts, and allow for easier interpretation by fellow healthcare professionals. Current methods for AD are often based on feature engineering, require large amounts training data, and have poor generalizability to novel settings. Secondly, supervised learning techniques generally require training data which can be expensive to obtain since health-care professionals must manually label it. This is not practical when building a model that needs to disambiguate many abbreviations. The goal of my research project is to develop a model that is trained to disambiguate over 80,000 medical abbreviations in a semi-supervised manner without the need for any manual labelling. I will generate a training set using a method that finds long forms in a corpus of clinical notes and replaces them with their corresponding abbreviation. To ensure all long forms are well represented, I will generate additional training samples by gathering sentences of words that are close to a given long form in vector space. Once I generate my training set, I will train a feed-forward neural network and test generalizability on a novel dataset. The outcomes of this project can be used in existing electronic medical records for abbreviation disambiguation tasks or in real-time by clinicians for on-the-spot abbreviation expansion to reduce note ambiguity.
Department
Program
Master's Award: Canada Graduate Scholarships
Recipient
Location
Ontario, Canada
Recipient Type
Academia
Type
Grant
Source
CIHR Funding Decisions Database
As published
Id
406450
Name
Skreta Marta
Region
Ontario
Country
Canada
Orgtype
Institution Paid; Research Institution
Prcname2
Special Cases - Awards Programs
Projecttitle
Improving word sense disambiguation in clinical narratives using deep learning
Programtype2
Studentship Programs
Orgnameinp2
University of Toronto
Cihramount2
$17,500
Cihrequipment2
$0
Programname2
Master's Award: Canada Graduate Scholarships
Keyworddelim
Abbreviation Disambiguation In Clinical Narratives; Deep Learning; Medical Notes; Medical Text Processing; Natural Language Processing; Semi-Supervised Learning
Abstract
My proposal is focused on using deep learning to disambiguate abbreviations in clinical notes so that they are more comprehensive. Abbreviation disambiguation (AD) is challenging since abbreviations can have many senses. For example, ra can mean right atrium, rheumatoid arthritis or room air depending on its context. Correctly expanding medical terms is important to build accurate patient profiles, link notes to ontological concepts, and allow for easier interpretation by fellow healthcare professionals. Current methods for AD are often based on feature engineering, require large amounts training data, and have poor generalizability to novel settings. Secondly, supervised learning techniques generally require training data which can be expensive to obtain since health-care professionals must manually label it. This is not practical when building a model that needs to disambiguate many abbreviations. The goal of my research project is to develop a model that is trained to disambiguate over 80,000 medical abbreviations in a semi-supervised manner without the need for any manual labelling. I will generate a training set using a method that finds long forms in a corpus of clinical notes and replaces them with their corresponding abbreviation. To ensure all long forms are well represented, I will generate additional training samples by gathering sentences of words that are close to a given long form in vector space. Once I generate my training set, I will train a feed-forward neural network and test generalizability on a novel dataset. The outcomes of this project can be used in existing electronic medical records for abbreviation disambiguation tasks or in real-time by clinicians for on-the-spot abbreviation expansion to reduce note ambiguity.
Competitiondate
201812
Approvedterm2
1 yr 0 mth
Pinamesdelim
Skreta, Marta
Approvedterm
1 an 0 mois
Cihrcontribution2
~ $17,500~ $0~