Large language models in radiology workflows: OpenEvidence dotflows as an accessible, evidence-centered resource.
Authors
Affiliations (4)
Affiliations (4)
- University of Toledo College of Medicine and Life Sciences, Toledo, OH, United States.
- Kansas City University of Medicine and Biosciences, Kansas City, MO, United States.
- Department of Radiology, University Hospitals Cleveland Medical Center, Cleveland, OH, United States. Electronic address: [email protected].
- Department of Radiology, University Hospitals Cleveland Medical Center, Cleveland, OH, United States.
Abstract
As large language models (LLMs) become increasingly integrated into radiology workflows, radiologists need to know which tools add value for which tasks. OpenEvidence (OE) is a literature-grounded clinical LLM platform, free to verified U.S. clinicians, that recently introduced "dotflows," customizable and shareable instruction templates designed to streamline clinical tasks. In this review, we propose that the distinctive value of OE dotflows lies in tasks that depend on retrieval of peer-reviewed literature and guidelines. We illustrate two such applications with examples from our own use, examination protocoling anchored to the American College of Radiology Appropriateness Criteria and case-based clinical teaching, including what worked and what required correction, and briefly address guideline-concordant recommendations in impressions. We distinguish these from tasks that restructure text the radiologist has already supplied, such as converting free dictation into a structured report, drafting an impression from supplied findings, or checking a report for internal inconsistencies, for which dotflows mainly offer the convenience of reusable instructions. We also describe how dotflow instructions can be drafted with a separate general-purpose LLM and outline the steps an institution must take before protected health information can be entered under a business associate agreement. All AI-generated content requires radiologist validation, and future studies should compare dotflow outputs with those of general-purpose LLMs given identical instructions and inputs.