Thursday, December 3 | 9:30 a.m.-10:30 a.m. | R3-SSIN07-7 | Room E450B
A two-stage large language model (LLM) pipeline can turn radiology reports into patient-friendly language while catching most of the errors that a single LLM pass leaves behind, according to this scientific presentation.
Radiology reports are often written at a college reading level, far above the sixth- to eighth-grade target recommended for patient materials, noted presenter Avery Knapp, Jr, MD, of teleradiology practice Expert Radiology, and colleagues. LLMs can simplify text, but they may omit or alter clinically important findings.
Using the MR-RATE dataset, the group had GPT-4o generate patient-friendly versions of more than 10,000 brain MRI and cervical spine reports. In a separate experiment, reports flagged as imperfect were reprocessed by a newer model, GPT-5.1.
GPT-4o lowered reading levels by more than five grades and preserved about 93% to 95% of findings. However, omissions, alterations, and fabrications still turned up in up to 4.5% of reports. Adding the second stage raised per-report content preservation to about 98%, recovering roughly three-quarters of omissions and two-thirds of alterations. Average readability remained at or below an eighth-grade level, according to the group.
“A two-stage LLM pipeline can deliver patient-readable radiology reports that meet plain-language standards while reducing the risk of clinically meaningful omissions,” the researchers] wrote.
Find out more at this Thursday morning talk.



