Abstract
Background: Large language models (LLMs) like ChatGPT are increasingly being explored for medical applications. However, their reliability in providing medication advice for patients with complex clinical situations, particularly those with multiple comorbidities, remains uncertain and under-investigated. This study aimed to systematically evaluate the performance, consistency, and safety of ChatGPT in generating medication recommendations for complex cardiovascular disease (CVD) scenarios. Methods: In this simulation-based study (21 January-1 February 2024), ChatGPT 3.5 and 4.0 were prompted 10 times for each of 25 scenarios, representing five common CVDs paired with five major comorbidities. A panel of five cardiologists independently classified each unique drug recommendation as "high priority" or "low priority". Key metrics included physician approval rates, the proportion of high-priority recommendations, response consistency (Jaccard similarity index), and error pattern analysis. Statistical comparisons were made using Z-tests, chi-square tests, and Wilcoxon Signed-Rank tests. Results: The overall physician approval rate for GPT-4 (86.90%) was modestly but significantly higher than that for GPT-3.5 (85.06%; p = 0.0476) based on aggregated data. However, a more rigorous paired-scenario analysis of high-priority recommendations revealed no statistically significant difference between the models (p = 0.407), indicating the advantage is not systematic. A chi-square test confirmed significant differences in error patterns (p < 0.001); notably, GPT-4 more frequently recommended contraindicated drugs in high-risk scenarios. Inter-model consistency was low (mean Jaccard index = 0.42), showing the models often provide different advice. Conclusions: While demonstrating high overall physician approval rates, current LLMs exhibit inconsistent performance and pose significant safety risks when providing medication advice for complex CVD cases. Their reliability does not yet meet the standards for autonomous clinical application. Future work must focus on leveraging real-world data for validation and developing domain-specific, fine-tuned models to enhance safety and accuracy. Until then, vigilant professional oversight is indispensable.