OBJECTIVES: Literature reviews have many applications in health economic and outcomes research, but are limited in breadth and depth by the time spent by reviewers, and are prone to human error and biases. Implementation of natural language processing (NLP) aims to address these issues. We reviewed the use of NLP in literature reviews, assessed the benefits and detriments of this, conducted our own test use case, and developed recommendations for future researchers.

METHODS: To identify use cases and information on NLP in literature reviews, we searched medical literature databases (PubMed, Science Direct, Google Scholar), conference abstracts lists, and other grey literature. The identified relevant studies are summarized herein. We further implemented NLP to conduct screening. Experts in systematic literature review were then consulted regarding the application of NLP to established literature review processes.

RESULTS: When used to conduct targeted literature reviews, NLP can reduce human labor, increasing breadth and depth at reduced costs. For systematic reviews, NLP can be employed to design and conduct searches, screen captured records, extract relevant information, and summarize this into key messages. However, these steps are not commonly implemented end-to-end fashion. NLP-based screening performs inconsistently; screening decisions may conflict with those of humans, with match rates varying from 51% to 96% in the literature. NLP literature review methods often do not comply with guidelines of the Preferred Reporting Items for Systematic Reviews and Meta-Analysis organization, limiting their applicability. We overcame the deficits of previous NLP-based reviews by reporting reasons for exclusion and attaining 100% match rate between human and NLP reviewers.

CONCLUSIONS: NLP can improve the breadth and depth of literature reviews while alleviating human labor and risk of bias/error. Literature reviewers should implement NLP with caution, giving precise instructions and sufficient training, verifying NLP decisions, and following practice guidelines where possible.