What’s Going On With Me and How Can I Better Manage My Health? The Potential of GPT-4 to Transform Discharge Letters Into Patient-Centered Letters to Enhance Patient Safety: Prospective, Exploratory Study (Preprint)

• DOI: 10.2196/preprints.67143
• Publication Date: 2024-10-07T19:27:11Z
• Authors: Felix Eisinger; Friederike Holder...; Moritz Mahling; Christian Stegema...; Anne Herrmann–Werner; Eric Nazarenus; Alessandra Sonanini; Martina Guthoff; Carsten Eickhoff; Martin Holderried
• Keywords: Jargon

<sec> <title>BACKGROUND</title> For hospitalized patients, the discharge letter is an important source of medical information, containing numerous instructions and health care tasks for patients to manage their own health. However, it usually written in professional jargon that inaccessible with little knowledge. Large language models such as GPT have potential translate summaries into patient-friendly letters. </sec> <title>OBJECTIVE</title> In this study, we used GPT-4 transform letters more readable patient evaluated how comprehensively safety-relevant information was identified transferred from patient-centered <title>METHODS</title> We developed three based on common conditions 72 defined “learning objectives.” Then, prompted The were analyzed quality, patient-centricity, identify learning objectives. Bloom’s taxonomy analyze categorize <title>RESULTS</title> While addressed majority (56/72; 78%) objectives letters, 11 (15%) not included A qualitative analysis showed Bloom category Understand (9/11) frequently missed than those Remember (2/11). Most missing pertained content field “prevention complications,” while regarding “lifestyle” “organizational” aspects mentioned often. Medical errors occurred a few (31/787; 4%) sentences. Regarding better readability using fewer terms (132/860; 15%) (165/273; 60%), well abbreviations (43/860; 5%) versus (49/273; 18%) explanations (121/131; 92%) (0/165; 0%). <title>CONCLUSIONS</title> Conclusion: Our study shows has patient-centricity are already well-established, do address all leading omission some aspects. Further optimization prompt-engineering might help overcome issue.