A Cross-Sectional Study of GPT-4–Based Plain Language Translation of Clinical Notes to Improve Patient Comprehension of Disease Course and Management

BackgroundThe 21st Century Cures Act provides patients with access to their clinical notes, but most patients, particularly those with limited health literacy, have difficulty understanding and utilizing them for health decisions and care management. Thus, there is a critical need to improve the readability of clinical notes to ensure that patient access to medical information facilitates equitable health outcomes. This study introduces a novel clinical application of Generative Pretrained Transformer 4 (GPT-4), assessing the effect of GPT-4–based plain language translation of discharge summary notes (DSNs) on subjective and objective comprehension, self-reported confidence, and time spent reading each DSN.MethodsWe enrolled 553 patients from December 2023 to February 2024. Participants were at least 18 years of age, able to read English, had no diagnosis of cognitive impairment, and had an appointment scheduled within 2 months of enrollment at the Duke University Health System, a multicenter academic health system of three inpatient hospitals and multiple outpatient clinics. Patients read four DSNs related to common reasons for admission to an inpatient general medical service: congestive heart failure, community-acquired pneumonia, diabetic ketoacidosis, and acute ischemic stroke. Two DSNs (selected at random) were replaced by GPT-4–based plain language translations. After reading each DSN, patients answered questionnaire items assessing subjective and objective comprehension. The effects of translation on comprehension outcomes were analyzed with linear mixed models.ResultsAcross all four DSNs, GPT-4–based translation improved objective comprehension by 1.2 of 4 points (CI, 1.09 to 1.26; P<0.001), subjective comprehension by 2.4 of 16 points (CI, 2.15 to 2.71; P<0.001), and self-reported confidence by 2.0 of 8 points (CI, 1.82 to 2.13; P<0.001). Improvements were greater among Black and Hispanic patients, older patients, male patients, and those who reported limited health knowledge. In particular, objective comprehension improved an additional 0.35 and 0.81 out of 4 points among Black and Hispanic patients, respectively.ConclusionsGPT-4–based translation substantially improved patient comprehension of DSNs, especially in populations that historically have low health literacy. Further research is needed to validate results in additional demographic groups and study downstream effects on health decision-making and outcomes.