Radiology report quality improvement through structured provider feedback: a data-driven initiative.
Authors
Affiliations (3)
Affiliations (3)
- Department of Radiology, University of Washington Medical Center, Seattle, USA. [email protected].
- Department of Radiology, University of Iowa Health Care, Iowa City, USA.
- Department of Radiology, University of Washington Medical Center, Seattle, USA.
Abstract
To evaluate EHR-embedded provider feedback workflow for identifying probable concerns in radiology reports and to assess large language model (LLM)-assisted classification of free text comments. We collected reports over 11 months (December 2024-October 2025) in this single-centered quality improvement initiative. Link was embedded in EHR (Epic), which directed providers to a voluntary REDCap form. Form captured satisfaction, report adequacy, follow-up requests, and free-text comments. Analysis included free text issue categorizations with Llama 3.3, for five pre-defined non-mutually exclusive categories, and sentiment analysis. Agreement with two radiologists was assessed using Cohen's kappa. Descriptive statistics were reported. Of 270 responses, 267 were included in the final analysis. Corresponded to 0.026% of 1,010, 219 final reports generated. Overall, 52.8% (141/267) of responses were rated ≥ 4 stars, 44.6% (119/267) indicated that the reports "somewhat" or "did not" address the clinical question, and 28.1% requested a follow-up in this voluntary cohort. Free-text comments were provided in 65.9% (176/267) of responses. 162 comments had a classifiable concern, of 176 free-text comments. The most frequent issue categories were clinical clarity and completeness (91.4%, 148/162), followed by actionability and recommendations (30.9%, 50/162). Kappa agreement showed moderate to substantial agreement between reviewers and LLM (0.607-0.734). Structured provider feedback can complement traditional peer review by capturing deficiencies that may go unnoticed by radiologists to radiologists' review alone. Human - LLM agreement was moderate to substantial, warranting further evaluation for scalable model.