Simplifying Radiology Reports for Patients Using ChatGPT : A Systematic Review

This systematic review investigated the benefits and risks of using ChatGPT to simplify radiology reports to enhance patient understanding. PubMed/MEDLINE, Embase, Web of Science, and CENTRAL databases were searched for studies published until December 17, 2024. In total, 16 studies met eligibility criteria for inclusion, where ChatGPT utilised the GPT-4 model in 10 studies (63%), GPT-4o in 1 study (6%), and GPT-3.5 in 5 studies (31%). Of the studies which simplified English radiology reports, the median reading grade level of the original reports ranged from 9.6 to 14 and post-simplification was 5-9.4, with factual accuracy that ranged from 78% to 100%, report completeness of 83% to 100%, and the incidence of potentially harmful errors was 0% to 36%, but this was reduced to 0% to 8% after excluding outlier studies. Only three studies (19%) evaluated comprehensibility according to non-medical laypersons. Overall, ChatGPT demonstrates substantial potential to assist in generating patient-friendly radiology reports but requires radiologist oversight, use of proprietary models, and compliance with local privacy legislations. Further prospective studies are required prior to widespread clinical use, including evaluation of patient comprehensibility and satisfaction, and impacts on radiologist workflow.