JMIR Publications

ABSTRACT

Background:

Literature Reviews (LRs) identify, evaluate, and synthesise relevant papers to a particular research question to advance understanding and support decision making. However, LRs, especially traditional systematic reviews are slow, resource intensive, and are outdated quickly.

Objective:

Using recent Natural Language Processing (NLP) and Unsupervised Machine Learning (UML) methods, this paper presents a tool named LiteRev that supports researchers in conducting LRs.

Methods:

Based on the user’s query, LiteRev can perform an automated search on different open-access databases and retrieve relevant metadata on the resulting papers. Papers (abstracts or full texts) are text processed and represented as a Term Frequency-Inverse Document Frequency (TF-IDF) matrix. Using dimensionality reduction (PaCMAP) and clustering (HDBSCAN) techniques, the corpus is divided into different topics described by a list of keywords. The user can select one or several topics of interest, enter additional keywords to refine their search, or provide key papers to the research question. Based on these inputs, LiteRev performs an iterative nearest neighbours search, and suggests a list of potentially interesting papers. The user can tag the relevant ones and trigger a new search until no additional paper is suggested for screening. To assess the performance of LiteRev, we ran it in parallel to a manual LR on the burden and care for acute and early HIV infection in sub-Saharan Africa. We assessed the performance of LiteRev using True and False Predictive Values, recall and Work Saved over Sampling.

Results:

We extracted, text processed and represented into a TF-IDF matrix 631 unique papers from PubMed. The topic modelling module identified 5 main topics and 16 topics (ranging from 13 to 98 papers) and extracted the 10 most important keywords for each. Then, based on 18 key papers, we were able to identify 2 topics of interest with 7 key papers in each of them. Finally, we ran the k-nearest neighbours module and LiteRev suggested first a list of 110 papers for screening, among which 45 papers were confirmed as relevant. From these 45 papers, LiteRev suggested 26 additional papers, out of which 8 were confirmed as relevant. At the end of the iterative process (4 iterations), 193 papers out of 613 papers in total (31.5% of the whole corpus) were suggested by LiteRev. After title/abstract screening, LiteRev identified 64 out of the 87 relevant papers (i.e., recall of 73.6%). After full text screening, LiteRev identified 42 out of the 48 relevant papers (i.e., recall of 87.5%, and Work Saved over Sampling of 56.0%).

Conclusions:

We presented LiteRev, an automation tool that uses NLP and UML methods to streamline and accelerate LRs and to support researchers in getting quick and in-depth overviews on any topic of interest.