The research behind eviStreams.

Read and cite the paper

The paper describes human oversight at three stages: approving an extraction plan, refining typed field definitions through a pilot, and reviewing proposed values before adjudication. Its evaluation spans four clinical corpora and three model families. A reported finding is that field specification has a stronger influence on extraction quality than model choice.

These findings describe the evaluation in the linked publication. They are not a guarantee of accuracy for new studies or a validation of every feature added to the live platform since publication.

Evaluate the workflow on your own review

Start with a pilot that reflects your study designs, reporting styles, and extraction fields. Compare proposed answers with independently checked references and inspect the source evidence behind disagreements.

• Record the field definitions, extraction configuration, model, and date used for the run.
• Define how missing information, multiple time points, units, and partially correct answers are scored.
• Report the number of studies and fields evaluated, including failures and excluded documents.
• Separate AI proposals from values corrected during human review.

This makes an evaluation interpretable and helps the team decide which fields need clearer instructions or closer review.

Research context

eviStreams was developed within Penn Dental Medicine’s Center for Integrative Global Oral Health. The platform connects structured extraction with independent review, adjudication, risk-of-bias assessment, and evidence synthesis.

Learn about the project or contact the team about research collaborations.