Proposed rubric flags unsafe AI-generated patient radiology reports

A rubric for evaluating AI-generated patient-friendly radiology reports was developed. When an AI model applied it to 480 reports, distribution decisions agreed with reference-standard decisions 88.1% of the time, suggesting a quality-assurance safeguard.

• The rubric scores five dimensions (clarity, content, certainty, tone, verbosity) on a three-point scale; reports receiving the lowest grade on any non-verbosity attribute are considered unsafe and should be withheld.
• In lay-participant testing of 480 reports, subjective judgments aligned with the rubric's distribution rule 73.5% of the time, while an AI model's distribution decisions matched reference grades 88.1% of the time.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.