EUVIMEDEUVIMED entdecken
Querschnittstudie

Comprehensive comparison of Turkish hypertension patient education texts generated by large language models: a cross-sectional content analysis

Rasim Gökmen, Kevser Sak, Batuhan Akın

Journal of Cardiology & Cardiovascular Surgery · 08.06.2026 · 4 (2) · 28-35

DOI: 10.51271/jccvs-0073 ↗

Kurz gefasst

Querschnittstudie in der Zeitschrift Journal of Cardiology & Cardiovascular Surgery, online erschienen am 08.06.2026. Verfasst von Rasim Gökmen und 2 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Aims

This study aimed to compare Turkish patient education texts on hypertension generated by large language models in terms of quality, understandability, actionability, readability, guideline adherence, and patient safety.

Methods

In a cross-sectional descriptive content analysis design, standard questions covering eight core domains of hypertension patient education were posed to ChatGPT, Gemini, and Claude using the same Turkish prompt. For each model, only the first response from a separate new chat session was analyzed, yielding a total of 24 texts. The quality of responses containing treatment or self-management information was evaluated using DISCERN, while the understandability and actionability of all texts were assessed with PEMAT-P, and readability was evaluated using the Ateşman formula. In addition, adherence to the 2025 AHA/ACC hypertension guideline was scored, and patient safety was classified dichotomously. Friedman, Wilcoxon, and Cochran’s Q tests were used for statistical analysis.

Results

No significant difference was found among the models in total DISCERN scores (p=0.066). There was an overall difference in the DISCERN overall quality score (p=0.022), but no significant pairwise difference was identified in post hoc comparisons. PEMAT-P understandability scores differed significantly among the models (p=0.004); Gemini and Claude scored higher than ChatGPT, and the ChatGPT–Gemini difference remained significant in post hoc analysis (p=0.012). PEMAT-P actionability scores were similar across models (p=0.472). For Ateşman readability scores, Claude scored higher than both ChatGPT and Gemini (p=0.030; p=0.017 for both pairwise comparisons). No significant difference was found in guideline adherence (p=0.717) or patient safety (p=0.135).

Conclusion

Large language models may be useful for generating Turkish hypertension patient education materials; however, their performance varies according to the evaluation domain. While differences were observed among models in understandability and readability, no clear superiority was demonstrated in guideline adherence or patient safety. Therefore, LLM-generated outputs should be used under physician supervision, particularly with respect to clinical content and safety.

Zitieren

APA

Gökmen, R., Sak, K., & Akın, B. (2026). Comprehensive comparison of Turkish hypertension patient education texts generated by large language models: a cross-sectional content analysis. Journal of Cardiology & Cardiovascular Surgery, 4(2), 28-35. https://doi.org/10.51271/jccvs-0073

Vancouver

Gökmen R, Sak K, Akın B. Comprehensive comparison of Turkish hypertension patient education texts generated by large language models: a cross-sectional content analysis. Journal of Cardiology & Cardiovascular Surgery. 2026;4(2):28-35. doi:10.51271/jccvs-0073

BibTeX
@article{Gkmen2026,
  author = {Gökmen, Rasim and Sak, Kevser and Akın, Batuhan},
  title = {Comprehensive comparison of Turkish hypertension patient education texts generated by large language models: a cross-sectional content analysis},
  journal = {Journal of Cardiology & Cardiovascular Surgery},
  year = {2026},
  volume = {4},
  number = {2},
  pages = {28--35},
  doi = {10.51271/jccvs-0073}
}
Kostenlos starten →