Accelerating Systematic Reviews Via Natural Language Processing

An essential part of developing practice guidelines and health policy involves examining the evidence generated by studies published that compare the effect of competing interventions on specific health outcomes. This is achieved through systematic reviews (SRs), which are rigorous reviews of several medical databases, on papers related to the topic of interest, within a specified time interval. However, the process of conducting SRs is labor-intensive and very time-consuming. In this paper, we apply natural language processing (NLP) techniques to enhance the efficiency of selecting relevant medical literature, thereby reducing the time required for researchers to manually screen, select, and review articles and identify relevant ones. The proposed algorithm automates the selection process by comparing titles and abstracts of articles to predefined criteria. If eligible, the article then undergoes full-text screening; otherwise, it is excluded. In the second stage, the full text is scrutinized to extract content that aligns with the intervention(s) and outcome(s) of interest. The proposed automation tool was tested in an umbrella review (SR of systematic reviews) on Vitamin D and falls, demonstrating significant benefit in reducing the time and effort required by researchers. The number of articles to be reviewed was reduced from 17,776 to 188, achieving a 100% recall rate, thus ensuring all relevant articles were included, and a 99% specificity rate, thus successfully excluding most irrelevant articles. This approach also led to a substantial reduction in the screening time, a 13.75 time-gain factor, decreasing the estimated review time from months to years to an estimated of two and a half months.