Large language model helps radiology department boost patients’ report comprehension

One radiology department has significantly bolstered patients’ comprehension of their imaging results with the help of a large language model. 

Emory University in Atlanta has utilized OpenAI’s GPT-5 program to create a web application that better serves its customers. The app combines LLM-generated radiology report summaries, clickable terms and AI-generated videos, according to research published Monday in JACR. 

In a prospective study involving 100 patients, lead author Hanzhou Li, MD, and colleagues produced promising early results. LLM-generated content helped meaningfully improve patient comprehension, with 48% of those surveyed identifying the AI report summaries as most helpful. 

“This prospective study demonstrates that interactive, LLM-driven applications can significantly improve self-reported patient comprehension of complex radiology reports, emphasizing their potential to enhance patient-centered communication,” Li, a resident physician and incoming faculty at Emory, and co-authors concluded. “However, patients had reservations about clinician-edited, LLM-generated summaries, indicating that successful integration is contingent on professional oversight—an added workload that may limit scalable real-world implementation.”

Researchers processed radiology reports with GPT-5 using a custom prompt, adapted from a prior study. Radiologist members of the team reviewed the outputs to help remove any extraneous language or AI hallucinations. Patients participating in the study were emailed their login to access AI-augmented reports, which featured radiologist-edited LLM summaries, clickable medical terms with embedded educational content, and AI-generated explanatory videos. The latter were produced by passing edited summaries through a video generation application programming interface (D-ID) and streamed directly within the web app. 

After reviewing the reports, patients then completed a survey evaluating comprehension, perceived usefulness and overall attitudes toward LLM-generated summaries using a five-point scale. Participants who completed the questionnaire also received a $15 gift card. Li and colleagues conducted the study in summer 2025 at a hospital-based outpatient imaging floor before scheduled scans at a university hospital.

Report comprehension improved from a median of 4/5 before the study to 5/5 afterward. However, LLM summaries often required manual clinician edits, at an average of about 24.75 words removed per summary and 0.13 added. When patients were asked if they were comfortable with such AI-based summaries without any corresponding physician edits, most reported being only somewhat comfortable (28%) or very comfortable (26%).

“Although patients recognized the value of clinician-edited LLM summaries, many expressed reservations regarding unsupervised AI deployment, emphasizing the importance of transparency and professional oversight,” the authors noted. “AI-generated summaries are not a replacement for clinician-patient discussion but may serve as a preparatory tool that enhances meaningful engagement. Sustainable integration will require thoughtful alignment of oversight, regulatory compliance, and workflow considerations to ensure safe and scalable adoption.” 

Read more, including potential study limitations, in the Journal of the American College of Radiology.