Thursday 12 January 2017

Natural Language Processing Use in Radiology: A Systematic Review

Following document is important with respect to Electronic Health Record -

"Natural Language Processing Use in Radiology: A Systematic Review" by Dorothy A. Sippo, MD, MPH, CIIP, Johns Hopkins University School of Medicine; Daniel Rubin, MD, MS; Paul G. Nagy, PhD, FSIIM; Brandyn Lau, MPH presented in the Annual Meeting of Society for Imaging Informatics in Medicine, SiiM2015

Hypothesis: The performance of natural language processing used in radiology has improved over time. The types of applications of this technology in radiology have expanded over time.

Introduction: The Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 has spurred greater adoption of electronic health records (EHRs), with their use nearly doubling from 2010-2011 (1). With more clinical data available in an electronic format, the goal is to be able to leverage this data to assess the quality of care and guide future decision making. A significant proportion of EHR data is stored in a narrative, unstructured format. Natural language processing (NLP) employs a computer to extract meaningful information from human language. In the setting of the EHR, it is typically used to extract information from unstructured report text. A 2008 review of research on the extraction of information from text documents in EHRs revealed that performance of these systems has improved since a prior systematic review in 1995 (2).
Radiology is a medical subspecialty with a rich collection of text documents reporting the results of imaging examinations. Frequently, these documents have been stored in an electronic format over significant periods of time. They represent an archive that is ripe for information extraction to identify imaging findings and disease diagnoses. There are numerous descriptions in the literature to NLP being applied within radiology (3). A systematic review of the literature to investigate the use of NLP in radiology would be helpful to summarize the progress in the field and to identify gaps. The goal of this work is to evaluate the performance of NLP over time in radiology. We will identify the types of information being extracted from radiology reports and the clinical applications of this informatics tool. We will also address the computer science methods being used for NLP in radiology. From our review, we will identify gaps in functionality and opportunities for future work.

Methods: We will use a systematic approach to searching the literature to minimize the risk of bias in selecting articles for inclusion in this review. Searching the literature will involve identifying reference sources, formulating a search strategy for each source, and executing and documenting each search. For our searches of electronic databases, we will identify relevant medical subject heading terms. 

Sources Our comprehensive search will include electronic searching of peer-reviewed literature databases and grey-literature databases as well as hand-searching. We will run searches of the MEDLINE®, EMBASE®, Cochrane Library, Scopus, Cumulative Index to Nursing and Allied Health Literature (CINAHL), Web of Science, INSPEC, and Compendex databases through September 15, 2014. We will design search strategies specific to each database to enable the team to focus the available resources on articles that are most likely to be relevant to the key questions about the performance and application of NLP over time within radiology. We will develop a core strategy for MEDLINE®, accessed via PubMed, on the basis of an analysis of the relevant medical subject heading terms and text words of key articles identified a priori. The PubMed strategy will form the basis for the strategies developed for the other electronic databases.
Management of Literature Search With the assistance of the Johns Hopkins University Welch Medical Library, all references will be downloaded into ProCite® version 5.0.3 (ISI ResearchSoft, Carlsbad, CA) and de-duplicated prior to initiating the review. We will use this database to store full articles in portable document format (PDF) and to track the search results at the title review, abstract review, article inclusion/exclusion levels.
Title Review Two team members will independently reviewed all titles. For a title to be eliminated at this level, both reviewers must indicate that it is ineligible. If the first reviewer marks a title as eligible, it will be promoted to the next level, or if the two reviewers do not agree on the eligibility of an article, it will automatically promoted to the next level.
Abstract Review We will exclude an abstract at this level if the abstract does not apply to one of the key questions or for any of the following reasons: does not address NLP used in radiology, has no original data (e.g., letter to the editor, comment, systematic review), or is not in English. Abstracts will be promoted to the article review level if two reviewers agreed that the abstract could be applicable. Differences of opinion will be resolved by discussion between the two reviewers.
Article Review Full articles that were selected for review during the abstract review phase will undergo independent review by two members of the study team to determine whether they should be included in the full data abstraction. If both reviewers determine the articles have applicable information, the articles will be included in the data abstraction.
Data Abstraction We will sequentially review each article to abstract data from the final list of articles. For all articles, reviewers will extract information on general study characteristics, including: study design, location, clinical topic of interest, inclusion and exclusion criteria, description of the population under study, and description of the NLP applications. In this process, the primary reviewer will complete all relevant data abstraction forms. A second reviewer will check the first reviewer’s data abstraction forms for completeness and accuracy. We will form reviewer pairs to include personnel with both clinical and methodological expertise. We will resolve differences of opinion through consensus adjudication between the reviewers.

Results: We will summarize the different types of NLP approaches used on radiology reports and their reported performance. We will describe how NLP is being applied to radiology and assess if applications have expanded over time. We will present the types of computer science methods used for NLP in radiology and, where possible, categorize these methodologies. We will also identify gaps in functionality and applications of NLP as it relates to modern challenges of quality assurance, business intelligence, decision support, and scientific discovery.

Discussion: We will discuss types of applications of NLP in radiology and how this has the potential to enable knowledge discovery to inform future healthcare decision making. We will highlight NLP applications with superior performance and factors that may have contributed to their success. We will discuss how an understanding of NLP methods can inform future development of NLP applications within radiology.

Conclusion: A systematic review of the literature of the use of NLP in radiology demonstrates how its performance and scope of applications have evolved over time and suggests new opportunities for research.

References:

  1. DesRoches CM, Charles D, Furukawa MF, Joshi MS, Kralovec P, Mostashari F, Worzala C, Jha AK. Adoption Of Electronic Health Records Grows Rapidly, But Fewer Than Half Of US Hospitals Had At Least A Basic System In 2012. Health Aff (Millwood). 2013 Aug;32(8):1478-85.
  2. Meystre SM, Savova GK, Kipper-Schuler KC, Hurdle JF. Extracting information from textual documents in the electronic health record: a review of recent research. Yearb Med Inform. 2008:128-44.
  3. Friedman C. A broad-coverage natural language processing system. Proc AMIA Symp. 2000:270-4.
---
Durjoy Majumder, Ph.D
Secretary, SSBTR