JMIR Publications

ABSTRACT

Background:

For hospitalized patients, the discharge letter is an important source of medical information, containing numerous discharge instructions and health care tasks for the patients to manage their own health. However, it is usually written in professional jargon that is inaccessible to patients with little medical knowledge. Large language models such as GPT have the potential to translate discharge summaries into patient-friendly letters.

Objective:

In this study, we used GPT-4 to transform discharge letters into more readable patient letters and evaluated how comprehensively patient safety-relevant information was identified and transferred from the discharge letters into patient-centered letters.

Methods:

We developed three discharge letters based on common medical conditions with 72 patient safety-relevant information, defined as “learning objectives.” Then, we prompted GPT-4 to transform the discharge letters into patient-centered letters. The patient letters were analyzed for medical quality, patient-centricity, and the potential to identify and translate the learning objectives. Bloom’s taxonomy was used to analyze and categorize learning objectives.

Results:

While GPT-4 addressed the majority (56/72; 78%) of the learning objectives from the discharge letters, 11 of the 72 learning objectives (15%) were not included in the majority of the patient-centered letters. A qualitative analysis based on Bloom’s taxonomy showed that learning objectives of the Bloom category Understand (9/11) were more frequently missed than those of the Bloom category Remember (2/11). Most of the missing learning objectives pertained to the content field “prevention of complications,” while learning objectives regarding “lifestyle” and “organizational” aspects were mentioned more often. Medical errors occurred in a few (31/787; 4%) of the sentences. Regarding patient-centricity, the patient-centered letters showed better readability than the discharge letters, using fewer medical terms (132/860; 15%) than the discharge letters (165/273; 60%), as well as fewer abbreviations (43/860; 5%) versus (49/273; 18%) and more explanations of medical terms (121/131; 92%) versus (0/165; 0%).

Conclusions:

Conclusion: Our study shows that GPT-4 has the potential to transform discharge letters into more patient-centered letters. However, while readability and patient-centricity are already well-established, the patient-centered letters do not comprehensively address all patient safety-relevant information, leading to the omission of some important aspects. Further optimization in the prompt-engineering might help to overcome this issue.