Supporting Systematic Review Production with Article Similarity Network Visualization

Ohio State University, Columbus OH

Investigators

Linked publications & trials

Abstract

PROJECT SUMMARY Systematic reviews (SRs), or systematic reviews of literature, summarize evidence drawn from high quality studies, and are often the preferred source of evidence-based practice (EBP). However, conducting an SR is labor-intensive and time consuming, typically requiring several months to complete. It has been reported that more than ten thousands of SRs are needed to synthesize existing medical knowledge. An Article screening process is one of the most intensive and time consuming steps, which requires SR researchers to screen a large amount of references, ranging from hundreds to more than 10,000 articles, depending on the size of a SR. In the past 10 years, machine learning model training approaches24-29 were developed to accelerate the article selection process through automation. However, they are not widely used due to diffusion challenges.7,14 Major obstacles include 1) a training sample is required to generate the automation algorithm. If the training sample is biased, the article selection process will systematically fail; 2) the automation approach is not made available for non-computer science specialists, therefore SR researchers will not be able to ?fine-tune? the automation algorithm for particular conditions in various SR topics; 3) As there is no global automation algorithm, the generalizability is significantly limited; 4) It is difficult to assess the actual workload saved, while finding every relevant article is required in SR. We propose a new approach to provide views of article relationships in an article network. This is different from other bibliometric networks constructing citation, co-author, or co-occurrence networks. Article network is a simple and logical concept: visualizing article relationships and distribution based on articles' similarities in titles, abstracts, keywords, publication types, etc. SR researchers can also alter the article distribution by adjusting the similarities. This approach does not aim to suggest an end-point of the screening process. Rather, it provides a view of distribution for included, excluded, and undecided articles. In the proposed research, we will integrate advanced techniques to sparsify article networks with mixed sparsification methods, and improve the quality and efficiency of large network visualization layouts by constructing a multi-level network structure and advanced force model. We aim to provide approaches to sparsify and visualize article networks with more than 10,000 articles. Our approach is highly generalizable that it can be used for any health science topics. By viewing the article distribution, SR researchers will be able to screen a large amount of literature more efficiently. This approach can be integrated into current SR technologies and used directly by SR researchers. The success of this project can support SR production on any health science topics, and thus streamline their ultimate application in EBP paradigms.