A large language model (LLM) helped patients better understand their own imaging reports in a prospective study from Emory University in Atlanta. The team used OpenAI’s GPT-5 to build a web application that combines AI-generated report summaries, clickable terms, and explanatory videos — and self-reported patient comprehension rose measurably.
The radiology report is a technical document, written by and for physicians. For the patient, it is often a wall of jargon. The question Emory’s group set out to answer is straightforward: can an LLM translate that text into something understandable without distorting the clinical content? The early results, published in the Journal of the American College of Radiology (JACR), suggest it can — provided there is physician oversight.
The Emory study in numbers
The work, led by resident physician Hanzhou Li, MD, enrolled 100 patients seen in summer 2025 on a hospital-based outpatient imaging floor at a university hospital, before scheduled scans. Each participant received an emailed login to an AI-augmented version of their report and, after reviewing it, completed a questionnaire scoring comprehension, perceived usefulness, and attitudes toward the LLM-generated summaries on a five-point scale. Those who finished the survey received a $15 gift card.
The headline finding: report comprehension improved from a median of 4/5 before the study to 5/5 afterward. Among the features offered, 48% of patients identified the AI-generated summaries as the most helpful. “This prospective study demonstrates that interactive, LLM-driven applications can significantly improve self-reported patient comprehension of complex radiology reports,” the authors concluded, emphasizing the potential to enhance patient-centered communication.
How the application works
Reports were processed by GPT-5 using a custom prompt adapted from a prior study. The app delivered three layers of patient support: radiologist-edited LLM summaries, clickable medical terms with embedded educational content, and AI-generated explanatory videos. Those videos were produced by passing the edited summaries through a video-generation API (D-ID) and streaming them directly within the web app.
It is an approach that echoes trends we have followed before, such as interactive multimedia reporting, which enriches the report with images and navigable elements. The difference is the audience: here, the recipient is the patient, not the referring physician.
The three-layer design is deliberate. A plain summary answers “what does this mean?”; clickable terms let curious patients dig deeper without overwhelming everyone; and the short video adds a human-like narration that many patients find more reassuring than text alone. Together they turn a static PDF into something closer to a guided explanation.
Human oversight is still essential
The most important point for radiologists may be this: the LLM summaries required manual editing. On average, about 24.75 words were removed per summary and only 0.13 added — meaning the clinician’s job was mostly pruning, stripping out excess language and potential hallucinations. Radiologist team members reviewed the outputs precisely to keep model errors from reaching patients. That editing burden is a real signal, not a footnote: it quantifies how much polishing a current-generation model still needs before its output is safe to hand to a patient.
Patients themselves voiced reservations. Asked whether they would be comfortable with AI summaries without any 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.
Implications for patient communication
For radiology services, the study points to a concrete value path: use LLMs not to replace the report, but to generate a patient-centered layer of explanation. This connects to results-access rules that, in many countries, deliver the report to patients almost in real time — often before they meet the ordering physician. A clear summary can reduce anxiety and improve engagement.
In the workflow, the challenge is integrating this layer without creating yet another siloed system. Platforms that bring reporting closer to the workstation, such as native reporting inside the radiologist’s cockpit, show the direction: the patient explanation should originate from the same environment where the report is produced, not a parallel app. Where digital report access is growing fast, this kind of feature could see rapid adoption.
Outlook and limitations
The authors are cautious. This is a small-sample, self-report, single-center study — far from definitive validation. They stress that AI summaries do not replace the clinician-patient conversation, serving at best as a preparatory tool that makes that dialogue more productive.
There is also the cost of oversight: if every summary needs physician review, scalability is limited. Sustainable integration will hinge on aligning oversight, regulatory compliance, and workflow. Even so, the direction is promising — generative AI appears to have found, in patient communication, a use case that is low in clinical risk and high in perceived value.