Background: In Belgium, patients can access their medical letters, but these are typically complex, written in medical language, covering many factors: disease activity, therapy, multimorbidity and follow-up plans.On the other hand, understandable medical information, tailored to patients needs is beneficial for adherence and shared decision-making. According to literature, giving patients access to their medical letters in an understandable format will improve health literacy, lead to more efficient use of care, less health inequality, and better compliance. Using artificial intelligence (AI) could help in providing a patient-friendly version of the medical letter, without extra burden for the physician.

Objectives: We aimed to develop an AI based workflow that automatically generates two patient-friendly versions of clinical letters:

• a detailed line-by-line translation preserving medical accuracy

• a simplified summary pointing out key actions.

In addition, a glossary explaining the complex medical terms used is provided.

Both versions, together with a disclaimer and the glossary, are combined into a single “patient-friendly letter” and automatically integrated into the patients’ electronic health records, becoming accessible to them via CoZo, the Belgian health data platform that provides secure, centralized access to personal health information. We aimed to obtain AI generated translations understandable by 80% of the population (language level B1).

Methods: Large Language Model prompting (ChatGPT) was developed and optimised using anonymised cardiology and nephrology letters from 3 hospitals, using an iterative approach. After each iteration, feedback from physicians, patient representatives and language experts specialized in adapting texts for non-native speakers and low-literate audiences (from integration agency Atlas VZW) was obtained, and used to improve the prompt. For the patient representatives scores were obtained for patient-friendliness, understandability and accessibility; for physicians additional scores for translation completeness, correctness, and risk were also obtained. Further evaluation covered compliance with language B1 level, and with GDPR, medical device regulation and the AI-act. Next, Rheumatology was selected to test whether this approach could be generalized across specialties, without adapting the prompt. For this purpose 20 rheumatology letters were evaluated during a supplementary test round by 5 rheumatologists, and patient experts from ReumaNet, a Belgian patient organisation, who provided 75 reviews.

Results: During the first iterations, several issues were identified that needed to be addressed through prompt adaptation or adjustments in the document intelligence workflow. These included avoiding overreading of lab and functional test results, and preventing incorrect handling of medication lists by excluding them from the translation flow (safe-by-design approach). In addition, other context traps had to be mitigated, such as addressing the patient when the letter was originally intended for a physician, or interpreting “not mentioned” as “absent.” Translation-specific issues also arose, including missing translations or suboptimal/incorrect translations. Layout issues were similarly noted, as the combination of disclaimer, line-by-line translation, summary, and glossary required careful structuring to remain clear and navigable for patients. After resolving these issues, the AI-generated patient-friendly letters achieved very high comprehension scores, as assessed by both physicians and patient experts. By comparison, the original letters were rated at only 16% understandable by patient experts. Language accessibility was generally improved to a stable B1 level in the summary version. The prompting approach proved fully transferable from cardiology and nephrology to rheumatology, resulting in consistently high scores for patient-friendliness, readability, accessibility, accuracy, completeness, and safety as assessed by both clinicians and patients (Table). Although the summary version was generally easier to read and understand, some patients considered it overly simplified or even patronizing, highlighting a tension between readability and preservation of detailed medical information. Maintaining the line-by-line version alongside the summary largely addressed this issue. This detailed version provides full content, closely aligns with the original letter, and explains each medical term in context. Patients who reviewed the results reported that the translations improved their insight into their medical situation and provided a greater sense of safety, as they could better understand communication about their disease and health status.

Conclusions: AI-generated patient-friendly letters, produced through fine-tuned prompting, were rated as satisfactory in terms of patient-friendly communication, and as understandable, accessible, and safe by both physicians and patients. While the summary version entails only minimal loss of detail and precision, the risk of significant misinformation remains low, mitigated by a legal disclaimer, the availability of the line-by-line version, and the reference to the included original letter. In addition, the summary consistently achieved a B1 language level, and the approach demonstrated successful cross-specialty generalization, indicating broad potential applicability. By providing patients with clear and reliable access to their own health information, this workflow supports shared decision-making, strengthens patient engagement, and facilitates communication across the care continuum, without adding workload for the treating physician.

REFERENCES: NIL.

Acknowledgments: NIL.

Disclosure of Interests: Ilse Hoffman: None declared, Karlien Vevey: None declared, Charles Cuigniez Element61, Veerle De Pourcq: None declared, Annelie Willems: None declared, Sandy Reinenbergh: None declared, Mark Helbert: None declared.