A shared latent space matrix factorisation method for recommending new trial evidence for systematic review updates

Area of Science:

• Medical Informatics
• Clinical Epidemiology
• Health Services Research

Background:

• Clinical trial registries are crucial for tracking evidence production and identifying outdated systematic reviews.
• Current methods for identifying relevant trials for review updates are labor-intensive, limiting their use.
• Automating this process is essential for efficient evidence synthesis.

Purpose of the Study:

• To evaluate a novel matrix factorization method for partially automating the identification of relevant clinical trial registrations for systematic review updates.
• To compare the performance of this new method against traditional document similarity approaches.

Main Methods:

• 179 systematic reviews on type 2 diabetes drug interventions were analyzed, with 537 associated ClinicalTrials.gov registrations.
• Trial registration text was used as features, transformed via Latent Dirichlet Allocation (LDA) or Principal Component Analysis (PCA).
• A matrix factorization approach using a shared latent space was developed and compared to document similarity for ranking relevant trials.

Main Results:

• The matrix factorization approach demonstrated superior performance, achieving a median rank of 59 and recall@100 of 60.9% with LDA features.
• This outperformed the document similarity baseline, which had a median rank of 138 and recall@100 of 42.8%.
• On an independent dataset, the best performing approach (document similarity) yielded a median rank of 67 and recall@100 of 62.9%.

Conclusions:

• A shared latent space matrix factorization method effectively ranks trial registrations, significantly reducing the manual workload for systematic review updates.
• This approach shows promise for integration into semi-automated pipelines to monitor new evidence for reviews.
• The findings support the development of computational tools to enhance the efficiency of evidence-based medicine.