Three Large Language Models (LLMs), One Heart: A Comparative Evaluation of ChatGPT, Claude, and DeepSeek in Cardiac...

Introduction Clear, accurate, and accessible patient education is central to informed decision-making and high-quality cardiovascular care. Large language models (LLMs) are increasingly being used to generate health information, offering the potential to rapidly produce patient education materials. However, concerns remain regarding the readability, quality, clinical accuracy, and adherence to evidence-based recommendations of AI-generated content. This study compared the performance of ChatGPT (OpenAI, San Francisco, CA), Claude (Anthropic PBC, San Francisco, CA), and DeepSeek (DeepSeek Artificial Intelligence Co., Ltd., Hangzhou, China) in generating patient education leaflets for three commonly performed cardiac imaging procedures. Methodology A cross-sectional study was conducted using standardized prompts to generate patient education leaflets for cardiac magnetic resonance imaging (CMR), coronary computed tomography angiography (CTCA), and invasive coronary angiography (ICA) using ChatGPT, Claude, and DeepSeek. Nine leaflets were evaluated for readability using Flesch-Kincaid Grade Level, Gunning Fog Index, Simple Measure of Gobbledygook (SMOG), Flesch Reading Ease, word count, and sentence count. Information quality was assessed using the modified DISCERN (mDISCERN) instrument. Guideline adherence was assessed using investigator-developed checklists based on recommendations from relevant professional societies and patient education resources. Expert clinical assessment was independently performed by two reviewers using a structured 16-point scoring system. Comparisons among the three models were performed using the Kruskal-Wallis test. Results DeepSeek demonstrated the most favorable overall readability profile, with the highest Flesch Reading Ease score (70.73 ± 1.74) compared with ChatGPT (50.70 ± 7.88) and Claude (60.93 ± 5.51; H(2) = 7.20, p = 0.027). DeepSeek achieved the highest mean mDISCERN score (3.83 ± 0.29), followed by ChatGPT (3.50 ± 0.50) and Claude (2.67 ± 0.29), although the difference was not statistically significant (H(2) = 5.593, p = 0.061). ChatGPT demonstrated the highest mean guideline adherence (91.54 ± 3.39%), followed by DeepSeek (89.33 ± 3.41%) and Claude (88.07 ± 6.48%; H(2) = 0.707, p = 0.702). DeepSeek achieved the highest expert clinical assessment score (16.00 ± 0.00), followed by ChatGPT (15.67 ± 0.58) and Claude (14.33 ± 0.58), with no statistically significant difference (H(2) = 5.394, p = 0.067). A significant difference was also observed in total word count (H(2) = 7.20, p = 0.027). Conclusions All three LLMs generated high-quality patient education leaflets for common cardiac imaging procedures, with each model demonstrating distinct strengths across different evaluation domains. DeepSeek showed the most favorable readability and achieved the highest information quality and expert clinical assessment scores, whereas ChatGPT demonstrated the greatest guideline adherence. However, all three models produced content above recommended patient health-literacy standards. LLMs may therefore serve as valuable clinician-assisted tools for developing cardiac imaging education materials, but expert review and readability optimization remain essential before clinical implementation.