Hospital discharge marks a critical transition where patients must understand complex instructions on medication, follow-up and self-care, yet standard discharge letters are often written for clinicians and can be difficult to interpret. Many patients struggle to retain key safety-related information, increasing the risk of errors after leaving hospital. To address this gap, an evaluation examined whether GPT-4–generated patient-friendly letters improve comprehension compared with conventional discharge documents.
How the Comparison Was Structured
The evaluation built on earlier work that had already shown GPT-4 could transform discharge letters into patient-directed versions that were more readable and more clearly structured, while also carrying some risk of omissions. Here, the emphasis moved from technical feasibility to comprehension. Each participant first read a brief case vignette with basic medical and socioeconomic information, then received either the standard discharge letter or the patient-friendly version, followed by the alternate version in a crossover design. The letters were presented digitally through a web-based interface, and participants could return to the text while answering questions, mirroring real access to written discharge information.
Must Read: Secure AI Improves Clinical Note Sectioning
For each of the three cases, 24 safety-related learning objectives were defined. These covered matters such as medication changes, warning symptoms and follow-up appointments. The objectives were evenly divided between Remember and Understand, then mapped to four content fields: Organization, Medication, Prevention of Complications and Lifestyle/Disease Management. Five objectives were excluded because they were not represented in the patient letters, leaving 67 for analysis. Responses were rated on a three-point scale ranging from not specified to fully specified. Two clinicians rated the answers independently, with very high agreement.
Overall Gains Favoured Patient Letters
Across all cases, the type of letter significantly affected comprehension. Patient letters led to higher rates of fully reported learning objectives and partially reported learning objectives, while reducing the proportion of omitted information. Fully stated objectives rose from 38.5% with standard discharge letters to 45.7% with patient letters. Partial reporting increased from 26.7% to 30%, and omissions fell from 34.8% to 24.3%. The overall odds ratio for better comprehension with patient letters was 1.74.
The effect extended across both Bloom levels. Participants performed better on Remember objectives than on Understand objectives regardless of letter type, with an odds ratio of 3.33 for Remember compared with Understand. Even so, patient letters improved performance in both categories. For Remember objectives, fully reported responses rose from 44.4% with discharge letters to 51% with patient letters. For Understand objectives, the increase was from 32.4% to 40.2%. No significant interaction was found between Bloom level and letter type, indicating that the advantage of patient letters remained consistent across both cognitive levels.
Medical correctness stayed high overall and did not differ significantly between the two letter types. Despite the gains, important gaps remained. Nearly a quarter of all learning objectives still went unreported after reading the patient letters. Six objectives were not reported in at least half of cases, and most of these belonged to the Understand category.
Medication and Organisation Showed the Strongest Gains
The benefit of patient letters varied by content field. The strongest improvement appeared in Medication. Fully reported medication objectives rose from 50.8% with standard discharge letters to 66.9% with patient letters. In the patient-friendly versions, medication changes and the purpose of each drug were stated explicitly in plain language, whereas the standard discharge letters conveyed the same information in a condensed table format without further explanation.
The second largest improvement appeared in Organization. Fully reported organisational objectives increased from 39.2% to 49.4%. These gains matter because the organisational domain covered follow-up tasks and appointments after discharge. The interaction between letter type and content field was statistically significant, showing that the advantage of patient letters depended on the type of information being conveyed.
Performance improved less in the more demanding domains. In Prevention of Complications, fully reported objectives increased only from 26.7% to 29.9%, although partially reported objectives rose from 22.5% to 32.8%. Even with the patient letters, this domain had the lowest rate of fully reported objectives, and 37.3% of key information in this field remained unknown. Lifestyle/Disease Management showed only a small increase in fully stated objectives, from 39.7% to 42.6%, alongside a slight decrease in partially reported objectives. These patterns suggest that plain language and clearer structure helped most with concrete and action-oriented content, while more interpretive or behaviour-related material remained harder to retain.
GPT-4–generated patient letters improved the comprehension of safety-relevant discharge information among standardised patients, with the clearest gains in medication-related and organisational content. They also improved performance across both lower-order and higher-order cognitive objectives, although Remember tasks remained easier than Understand tasks. At the same time, the results showed clear limits. Important information continued to go unrecognised, especially in Prevention of Complications and Lifestyle/Disease Management. The findings therefore point to a useful but incomplete communication support. Patient letters strengthened understanding of some of the most actionable parts of discharge information, but they did not resolve all comprehension gaps, particularly where content required more complex interpretation or broader self-management understanding.
Source: Journal of Medical Internet Research
Image Credit: iStock
References:
Holderried F, Sonanini A, Stegemann–Philipps C et al. (2026) Impact of GPT-4–Generated Discharge Letters on Patients’ Medical Comprehension: Prospective Crossover Study
J Med Internet Res;28:e81243.
