Croat Med J. 2026 Oct 31;67(5):329-337.
ABSTRACT
AIM: To compare ChatGPT's and Claude's responses to frequently asked patient questions about deep vein thrombosis (DVT) in terms of expert-rated accuracy, comprehensibility, and unnecessary detail.
METHODS: We compiled 33 Turkish questions about DVT from online sources and clinical experience. The questions were posed to ChatGPT and Claude in separate sessions, and the models' unedited responses were collected. Three blinded cardiovascular surgeons independently evaluated each answer on a 1-5 Likert scale across three domains: accuracy, comprehensibility, and unnecessary detail. We compared average answer lengths and domain scores using paired statistical tests.
RESULTS: The mean lengths of the responses did not differ significantly between the models (139.7±49.4 words vs 145.67±25.41 words; P=0.490). ChatGPT had a higher median rater-averaged comprehensibility score (4.33 vs 4.00; raw P=0.002). In the principal rater-averaged analyses, the models did not significantly differ in accuracy (Holm-adjusted P=0.085), whereas ChatGPT received significantly higher scores for comprehensibility and unnecessary detail (Holm-adjusted P=0.006 for both).
CONCLUSION: Our findings suggest potential differences in the presentation of patient-facing information rather than overall clinical superiority of any model. AI-generated medical information should continue to be reviewed by qualified health care professionals.
PMID:42963998