Automated mapping of electronic data capture fields to SDTM
Eric Yang1, Laura Katz1, Sushila Shenoy1
1Medidata Solutions, New York, New York, United States of America.
Plos One
|November 7, 2024
Summary
Siamese networks automate clinical trial data mapping, significantly improving accuracy and reducing manual work. This machine learning approach enhances data capture to submission efficiency.
Area of Science:
- Machine Learning
- Bioinformatics
- Clinical Data Management
Background:
- Manual data mapping in clinical trials is labor-intensive.
- Traditional methods lack efficiency for large datasets and numerous classes.
Purpose of the Study:
- To reduce manual effort in clinical trial data capture and regulatory submission.
- To demonstrate the efficacy of Siamese networks for data field classification.
Main Methods:
- Utilized Siamese networks to generate embeddings from electronic data capture form metadata.
- Employed traditional machine learning classifiers on generated embeddings.
- Focused on learning data similarity for improved classification.
Main Results:
- Achieved classification accuracies greater than 90%, a significant improvement over traditional methods.
- Demonstrated a 15% increase in accuracy in many cases.
- Identified lack of training data as a cause for some inaccuracies.
Conclusions:
- Siamese networks efficiently generate low-dimensional data field embeddings.
- Automated data schema mapping is achievable with high accuracy.
- This represents a key step towards automating clinical trial data management tasks.
