Bidirectional Encoder Representations from Transformers (BERT) and BERT-based approaches are the current state-of-the-art in many natural language processing (NLP) tasks; however, their application to document classification on long clinical texts is limited. In this work, we introduce four methods to scale BERT, which by default can only handle input sequences up to approximately 400 words long, to perform document classification on clinical texts several thousand words long. We compare these methods against two much simpler architectures - a word-level convolutional neural network and a hierarchical self-attention network - and show that BERT often cannot beat these simpler baselines when classifying MIMIC-III discharge summaries and SEER cancer pathology reports. In our analysis, we show that two key components of BERT - pretraining and WordPiece tokenization - may actually be inhibiting BERT's performance on clinical text classification tasks where the input document is several thousand words long and where correctly identifying labels may depend more on identifying a few key words or phrases rather than understanding the contextual meaning of sequences of text.
|Number of pages||12|
|Journal||IEEE Journal of Biomedical and Health Informatics|
|State||Published - Sep 2021|
Bibliographical noteFunding Information:
Manuscript received September 14, 2020; revised February 5, 2021; accepted February 22, 2021. Date of publication February 26, 2021; date of current version September 3, 2021. This work was supported in part by the Joint Design of Advanced Computing Solutions for Cancer (JDACS4C) program established by the U.S. Department of Energy (DOE) and National Cancer Institute (NCI) of the National Institutes of Health, in part by the auspices of the U.S. Department of Energy by Argonne National Laboratory under Contract No. DE-AC02-06-CH11357, in part by Lawrence Livermore National Laboratory under Contract No. DEAC52-07NA27344, in part by Los Alamos National Laboratory under Contract No. DE-AC5206NA25396, and in part by Oak Ridge National Laboratory under Contract No. DE-AC05-00OR22725. This work used resources of Oak Ridge Leadership Computing Facility at Oak Ridge National Laboratory, which was supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725. This work was supported by Exascale Computing Project (17-SC-20-SC), a collaborative effort of the U.S. Department of Energy Office of Science and National Nuclear Security Administration. (Corresponding authors: Shang Gao; Georgia Tourassi.) Shang Gao, Mohammed Alawad, M. Todd Young, John Gounley, Noah Schaefferkoetter, Hong Jun Yoon, and Georgia Tourassi are with Oak Ridge National Laboratory, Oak Ridge, TN 37830 USA (e-mail: email@example.com; firstname.lastname@example.org; email@example.com; firstname.lastname@example.org; email@example.com; firstname.lastname@example.org; email@example.com).
© 2013 IEEE.
- clinical text
- deep learning
- natural language processing
- neural networks
- text classification
ASJC Scopus subject areas
- Computer Science Applications
- Health Informatics
- Electrical and Electronic Engineering
- Health Information Management