Classifiers of Data Sharing Statements in Clinical Trial Records

Area of Science:

• Computational linguistics
• Clinical trial data management
• Bioinformatics

Background:

• Digital individual participant data (IPD) from clinical trials is valuable for scientific reuse.
• Identifying available IPD requires interpreting textual data-sharing statements (DSS) in large databases.
• Computational linguistics, particularly pre-trained language models, offers potential for automating this interpretation.

Purpose of the Study:

• To evaluate the effectiveness of domain-specific pre-trained language models in classifying textual data-sharing statements (DSS).
• To compare classifier performance in reproducing original availability categories versus manually annotated labels for IPD.
• To assess the potential of these classifiers in aiding the automatic identification of available IPD.

Main Methods:

• Utilized a subset of 5,000 textual DSS from ClinicalTrials.gov.
• Developed and evaluated classifiers based on domain-specific pre-trained language models.
• Compared classifier performance against original availability categories and manually annotated labels using standard metrics.

Main Results:

• Classifiers trained to predict manually annotated labels outperformed those trained on original availability categories.
• This indicates that textual DSS contain information not captured by the original availability categories.
• The findings suggest significant potential for automated IPD identification.

Conclusions:

• Domain-specific pre-trained language models can effectively classify textual data-sharing statements.
• Manual annotations provide richer information for IPD availability classification than original categories.
• These advanced classifiers can significantly aid in the automatic discovery of reusable individual participant data from large clinical trial databases.