Can Large Language Models Answer Patient Questions About Lymphedema and Lipedema? A Blinded Comparative Expert Evaluation
Turkish Journal of Plastic Surgery, cilt.35, sa.1, ss.31-40, 2026 (ESCI, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 35 Sayı: 1
- Basım Tarihi: 2026
- Doi Numarası: 10.4103/tjps.tjps_71_26
- Dergi Adı: Turkish Journal of Plastic Surgery
- Derginin Tarandığı İndeksler: Emerging Sources Citation Index (ESCI), Scopus, Directory of Open Access Journals, Academic Search Ultimate (EBSCO)
- Sayfa Sayıları: ss.31-40
- Anahtar Kelimeler: Artificial intelligence, consumer health information, health literacy, lipedema, lymphedema, patient education
- Erzincan Binali Yıldırım Üniversitesi Adresli: Evet
Özet
Background: – Artificial intelligence (AI) chatbots are increasingly used by patients to obtain health information about lymphedema and lipedema, two chronic and frequently overlapping conditions characterized by diagnostic delay and limited awareness. Objectives: – The objective of the study was to compare the quality, clinical utility, harm potential, and readability of responses generated by three widely used large language models to standardized patient questions about lymphedema and lipedema. Materials and Methods: – In this blinded, cross-sectional, in silico study, 28 patient-oriented questions (13 lymphedema and 15 lipedema) drawn verbatim from the Lymphatic Education & Research Network were submitted to ChatGPT, Gemini, and Claude, yielding 84 anonymized and randomized responses. A nine-member expert panel of plastic and reconstructive surgeons, cardiovascular surgeons, and physical medicine and rehabilitation specialists independently scored each response for clinical depth, technical accuracy, decision utility, a modified Global Quality Score, and harm potential. Readability was assessed with the Flesch Reading Ease Score and the Flesch–Kincaid Grade Level. Results: – Model performance was disease-dependent. For lymphedema, ChatGPT achieved significantly higher technical accuracy, decision utility, and overall quality than Gemini and Claude (all P < 0.001), whereas no significant inter-model differences were observed for lipedema. Harm scores did not differ significantly among models, and all models exceeded recommended patient reading levels, with lipedema responses being less readable than lymphedema responses. Conclusions: – AI chatbot performance varies by disease. No model reached the recommended sixth-grade reading level, and no accuracy, utility, or harm threshold was predefined, so adequacy for autonomous patient use cannot be judged; clinician oversight remains necessary.