OP129 The Use Of A Text-Mining Screening Tool For Systematic Review Of Treatments For Relapsed/Refractory Diffuse Large B-Cell Lymphoma

Abstract

Introduction: Human screening of title and abstracts in a systematic literature review (SLR) is labor intensive and time-consuming. In many instances, thousands of citations may be retrieved; the vast majority excluded upon screening. Text-mining semi-automates and accelerates screening by identifying patterns in relevant and irrelevant citations, as labelled by the screener. One such text-mining tool, Abstrackr, uses an algorithm within an active-learning framework to predict the likelihood of citations being relevant. The objective of this study was to assesses the performance of Abstrackr for title and abstract screening in an SLR of treatments for relapsed/refractory diffuse large B-cell lymphoma. Methods: Citations identified from searches of electronic databases were imported to Abstrackr. An investigator-selected database of terms indicating relevance of title and abstract to the research question were uploaded. These terms were partly informed by the SLR inclusion/exclusion criteria. Citations deemed most relevant by Abstrackr were screened first (screening prioritization). Screening was carried out until a maximum prediction score of 0.4 or less, based on previous experience in the literature, was reached. Remaining citations were deemed unlikely to be relevant and did not undergo screening (screening truncation). Separately, a single-human screener screened all citations using Covidence. Results: A total of 7,723 citations and 154 initial terms were uploaded to Abstrackr. Of these citations, 2,572 (33 percent) were screened before a prediction score of 0.39 was reached. Compared to single-human screening (conducted on all citations), the workload saving associated with Abstrackr was 5 days. A total of 451 (6 percent) citations proceeded to full-text screening; ten (0.1 percent) were included in the final evidence base. No citations predicted to be irrelevant by Abstrackr were included in the final evidence base. Conclusions: Text-mining tools such as Abstrackr have the potential to reduce workload associated with title and abstract screening, without missing relevant citations.