Automatically disambiguating medical acronyms with ontology-aware deep learning

Abstract

Modern machine learning (ML) technologies have great promise for automating diverse clinical and research workflows; however, training them requires extensive hand-labelled datasets. Disambiguating abbreviations is important for automated clinical note processing; however, broad deployment of ML for this task is restricted by the scarcity and imbalance of labeled training data. In this work we present a method that improves a model’s ability to generalize through novel data augmentation techniques that utilizes information from biomedical ontologies in the form of related medical concepts, as well as global context information within the medical note. We train our model on a public dataset (MIMIC III) and test its performance on automatically generated and hand-labelled datasets from different sources (MIMIC III, CASI, i2b2). Together, these techniques boost the accuracy of abbreviation disambiguation by up to 17% on hand-labeled data, without sacrificing performance on a held-out test set from MIMIC III.

PhenoPad: Building AI enabled note-taking interfaces for patient encounters

“… Medical term extraction: We used Neural Concept Recognizer (NCR) for medical term extraction from notes and speech 30 . Abbreviation disambiguation: Our abbreviation disambiguation model was described in and trained on public medical notes 34 . Clinical decision support: We applied the clinical decision support methods implemented in PhenoTips 35 .…”

“…Abbreviation disambiguation: Our abbreviation disambiguation model was described in and trained on public medical notes 34 .…”

Deciphering clinical abbreviations with a privacy protecting machine learning system

“…Key components include the separate tasks of detecting abbreviations in free text snippets and the expansion of those abbreviations into concepts or long forms. The number of abbreviations that prior methods have evaluated varies from 13 to 1116 15 – 18 , with separate models typically developed for each abbreviation. Many studies include only abbreviations that are “ambiguous” (i.e., multiple long forms for the abbreviation) 18 , although unambiguous abbreviations can be difficult even for physicians in other specialties to discern 19 , 20 .…”

“…The number of abbreviations that prior methods have evaluated varies from 13 to 1116 15 – 18 , with separate models typically developed for each abbreviation. Many studies include only abbreviations that are “ambiguous” (i.e., multiple long forms for the abbreviation) 18 , although unambiguous abbreviations can be difficult even for physicians in other specialties to discern 19 , 20 . To detect abbreviations in text, prior research focuses on heuristics, such as string matching of abbreviations like “ivf,” rather than machine learning 20 – 22 .…”

“…A variety of machine-learning approaches have been developed for disambiguating abbreviations in clinical text, including naive Bayes 23 , support vector machines 23 , 24 , profile-based approaches 25 , algorithms based on hyperdimensional computing 26 , convolutional neural networks 27 , long short-term memory networks 28 , 29 , encoder-based transformers (e.g. clinicalBERT) 18 , 30 , latent meaning cells 31 , and decoder-based transformers 32 . A recent study 18 introduced a model that predicts the correct expansion of a detected abbreviation from all its possible senses.…”