Large Language Model Simplification of Ultrasound Reports Improves Patient Understanding
Large Language Model Simplification of Ultrasound Reports Improves Patient Understanding
This multicenter, patient-blinded, controlled evaluation assessed whether expert-reviewed artificial intelligence (AI)-simplified ultrasound reports improved patient- or guardian-reported understanding and reading experience compared with standard ultrasound reports. Routine ultrasound reports were completed through existing clinical processes. After completion of the routine report, participants were assigned to view either the standard report or an expert-reviewed plain-language version generated with a large language model workflow.
The simplified report was intended only as a patient-facing communication aid. It did not replace the standard clinical report and did not alter ultrasound acquisition, diagnostic interpretation, treatment decisions, follow-up, or subsequent clinical management. Patient- or guardian-reported outcomes included cognitive workload, comprehension, report perception, and reading time. Expert review assessed whether AI-generated simplified reports preserved source meaning and identified factual errors, omissions, or unsupported additions before patient presentation.
Radiology and ultrasound reports are primarily written for communication among clinicians and often contain technical terminology, compressed syntax, measurements, uncertainty statements, and diagnostic language that may be difficult for patients to interpret. Patient access to reports through electronic portals may improve transparency, but access alone does not ensure comprehension. This study evaluated a bounded AI-assisted communication workflow designed to translate completed ultrasound reports into patient-facing plain language while preserving the meaning of the original clinical report.
The study was conducted at three participating centers: Sichuan Cancer Hospital and Institute, Sichuan Provincial Hospital of Integrated Traditional Chinese and Western Medicine, and The First People's Hospital of Liangshan Yi Autonomous Prefecture. Eligible participants were patients receiving an ultrasound report at a participating center. For participants younger than 18 years, a legal guardian provided consent, read the report, and completed the questionnaire.
Routine ultrasound reports were first completed through standard clinical processes. The study workflow began only after the routine report had been finalized. A built-in random-number function determined whether the participant or guardian viewed the standard report or an expert-reviewed AI-simplified report. Participants or guardians were not informed which presentation format they received. The standard clinical report remained the authoritative diagnostic report and the sole report used for clinical communication and care.
For the AI-simplified format, the completed routine report text was submitted to a large language model workflow. The prompt instructed the model to preserve all findings, avoid new diagnoses or recommendations, explain terminology in plain language, and maintain uncertainty expressed in the source report. Before presentation to participants or guardians, the generated simplified report was reviewed against the routine report by paired ultrasound physicians at each center. Experts corrected inaccuracies, omissions, or unsupported additions before patient exposure. The workflow did not generate orders, referrals, follow-up intervals, medication instructions, or treatment recommendations, and no model output was written back into the diagnostic report.
After reading the assigned report format, participants or guardians completed a questionnaire evaluating their immediate reading experience. Patient- or guardian-reported outcomes included cognitive workload, text comprehension, report perception, and approximate reading time. Cognitive workload included mental demand, frustration, and effort, with lower scores indicating a more favorable experience. Comprehension included clarity, readability, and comprehensibility. Report perception included helpfulness, informativeness, and trust, with higher scores indicating more favorable responses.
Expert assessment of AI-generated simplified reports included factual errors, important omissions, unsupported additions, terminology simplification, clinical usefulness, and overall quality. Paired expert ratings were resolved into consensus records for report-level safety summaries. The primary purpose of expert review was to determine whether the AI-generated patient-facing explanation preserved the source meaning and to define the residual need for human oversight before such reports are shown to patients.
The prespecified analysis population included completed evaluations with recorded center and presentation format. Outcomes were analyzed using descriptive statistics, between-format comparisons, and adjusted models including presentation format, age, sex, education, and center. The study was approved by the Medical Ethics Committee of Sichuan Cancer Hospital, and written informed consent was obtained from adult participants or from legal guardians of minors.
Inclusion Criteria:
Participants or legal guardians able to read one assigned report presentation format and complete the study questionnaire immediately after reading.
Exclusion Criteria:
Inability of the participant or legal guardian to read the assigned report presentation format or complete the questionnaire.
Incomplete evaluation, including missing responses to any of the nine rating items or the reading-time item.
Missing key study information required for analysis, including participating center or assigned report presentation format.