Quantitative Evaluation of an Agentic Workflow for Study-Publication Linkage
Abstract
Clinical trial evidence is fragmented across registries and publications, making it difficult to determine whether a publication belongs to an existing study or is the first of a new study. This thesis first constructs a Cochrane-derived benchmark dataset by extracting study entities, publication records, and study–publication links from systematic reviews. The resulting dataset supports large-scale evaluation of Study–Publication linkage and enables analysis across medical domains.
The thesis investigates whether a highly autonomous, tool-augmented agentic workflow with shared state can improve this linkage task. Given a publication record, the agent can independently invoke available tools and subagents to search, compare, verify, and explain candidate matches. The study asks: to what extent does this workflow improve precision, recall, and F1 for existing-study linkage, and BERTScore for new-study suggestions; how much do different tools and assistants contribute, measured through ablation studies; and whether the architecture generalizes across MeSH-derived domains such as mental health, cancer, cardiovascular diseases, and infections.
| Attribute | Value |
|---|---|
| Title (de) | Quantitative Evaluatierung eines Agentischen Workflows für Studien-Publikation Verknüpfung |
| Title (en) | Quantitative Evaluation of an Agentic Workflow for Study-Publication Linkage |
| Project | |
| Type | Master's Thesis |
| Status | started |
| Student | Jonathan Parth |
| Advisor | Joshua Oehms |
| Supervisor | Prof. Dr. Florian Matthes |
| Start Date | 19.06.2026 |
| Sebis Contributor Agreement signed on | 01.06.2026 |
| Checklist filled | No |
| Submission date | 21.12.2026 |