Patient-Friendly Discharge Summaries in Korea Based on ChatGPT: Software Development and Validation
10.3346/jkms.2024.39.e148
- Author:
Hanjae KIM
1
;
Hee Min JIN
;
Yoon Bin JUNG
;
Seng Chan YOU
Author Information
1. College of Nursing, Yonsei University, Seoul, Korea
- Publication Type:Original Article
- From:Journal of Korean Medical Science
2024;39(16):e148-
- CountryRepublic of Korea
- Language:EN
-
Abstract:
Background:Although discharge summaries in patient-friendly language can enhance patient comprehension and satisfaction, they can also increase medical staff workload. Using a large language model, we developed and validated software that generates a patient-friendly discharge summary.
Methods:We developed and tested the software using 100 discharge summary documents, 50 for patients with myocardial infarction and 50 for patients treated in the Department of General Surgery. For each document, three new summaries were generated using three different prompting methods (Zero-shot, One-shot, and Few-shot) and graded using a 5-point Likert Scale regarding factuality, comprehensiveness, usability, ease, and fluency. We compared the effects of different prompting methods and assessed the relationship between input length and output quality.
Results:The mean overall scores differed across prompting methods (4.19 ± 0.36 in Few-shot, 4.11 ± 0.36 in One-shot, and 3.73 ± 0.44 in Zero-shot; P < 0.001). Post-hoc analysis indicated that the scores were higher with Few-shot and One-shot prompts than in zero-shot prompts, whereas there was no significant difference between Few-shot and One-shot prompts. The overall proportion of outputs that scored ≥ 4 was 77.0% (95% confidence interval: 68.8–85.3%), 70.0% (95% confidence interval [CI], 61.0–79.0%), and 32.0% (95% CI, 22.9–41.1%) with Few-shot, One-shot, and Zero-shot prompts, respectively. The mean factuality score was 4.19 ± 0.60 with Few-shot, 4.20 ± 0.55 with One-shot, and 3.82 ± 0.57 with Zero-shot prompts. Input length and the overall score showed negative correlations in the Zero-shot (r = −0.437, P < 0.001) and One-shot (r = −0.327, P < 0.001) tests but not in the Few-shot (r = −0.050, P = 0.625) tests.
Conclusion:Large-language models utilizing Few-shot prompts generally produce acceptable discharge summaries without significant misinformation. Our research highlights the potential of such models in creating patient-friendly discharge summaries for Korean patients to support patient-centered care.