Description
- Abstract:
- Evidence synthesis products (e.g., systematic reviews, rapid reviews) form the basis of evidence-based healthcare. Literature identification—searching for studies and screening the resulting citations to identify relevant literature—is an important but time- and labor-intensive step in the systematic review process. Medical librarians serve an important role in developing search queries that balance the requirement to identify all relevant studies with the constraints of the team’s available time and budget, but this process is time-consuming, and not every review team has access to an experienced librarian. Thus, there is a clear need for computer-based tools that assist in the design and development of high-quality search queries. With the ongoing development of tools to semi-automate literature search query development, it is increasingly important to find meaningful ways to evaluate their performance. In the first of three sections of this dissertation, I describe a discrete choice experiment, conducted as a survey, to determine which types of measures a variety of stakeholders prefer. Surveyed systematic review methodologists and librarians would like to see studies report measures used to establish whether a tool identifies all relevant records as a ratio (i.e., sensitivity) and prefer measures that clearly indicate how much work it saves. In the second and third sections of this dissertation, I describe a project in which my colleagues and I divided a corpus of systematic review topics and search queries into training and validation sets. We subsequently developed an independent evaluation dataset. We trained language models on the training dataset and, after settling on all hyper-parameter settings using the validation dataset, evaluated them quantitatively, using the evaluation dataset. The models had a median sensitivity of 85% and required that the simulated team screen approximately 1,000 abstracts for every included citation. I also evaluated the models qualitatively, through semi-structured interviews with eight librarians, during which they piloted and evaluated the models on real search topics. The librarians generally expressed that although the tool-generated queries lacked both the necessary sensitivity and precision to be used without scrutiny, the queries could be used as teaching tools or as starting points for non-expert searchers.
- Notes:
- Thesis (Ph. D.)--Brown University, 2024
Citation
Adam, Gaelen Phyfe,
"Advancing the Semi-automation of Literature Identification for Evidence Synthesis: Evaluation Measure Prioritization, Training Data Development, and Language Model Implementation"
(2024).
Health Services, Policy & Practice Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://repository.library.brown.edu/studio/item/bdr:d56axgw9/
Relations
Collection:
-
Health Services, Policy & Practice Theses and Dissertations
Theses and Dissertations for the Health Services, Policy & Practice department....