Accuracy and Reliability of AI Models in Emergency Myocardial Infarction Education

Acute myocardial infarction (AMI) is a major global cause of morbidity and mortality. Large language models (LLMs) are emerging tools for patient education. This study evaluated the performance of three LLMs in delivering accurate, reliable, and readable educational content regarding AMI.

In this cross-sectional study (February-March 2025), a clinical case of a patient with an inferior STEMI ECG was presented to three LLMs: ChatGPT-4o, Claude 3.7 Sonnet, and Gemini Advanced 2.0 Flash. Each model answered 30 patient-focused questions across three domains: general disease knowledge, diagnostic processes, and treatment approaches. Responses were assessed by four emergency medicine associate professors (10-20 years of experience) using a 5-point Likert scale for accuracy, DISCERN and EQIP tools for reliability and quality, and standard readability indices.

ChatGPT-4o achieved the highest accuracy score (4.38 ± 0.38), followed by Claude 3.7 (4.09 ± 0.55) and Gemini 2.0 (3.92 ± 0.41) (p < 0.001). ChatGPT-4o performed significantly better in general information (p = 0.002) and diagnostics (p = 0.009), while Claude 3.7 excelled in treatment-related content (p = 0.015). Claude 3.7 produced significantly more readable responses than both ChatGPT-4o and Gemini 2.0 across all indices (Flesch-Kincaid, p = 0.002, Gunning Fog, p ≤ 0.001, Coleman-Liau, p = 0.003). ChatGPT-4o scored "excellent" on the DISCERN scale; all models were rated as "good quality with minor shortcomings" on EQIP. Reliability scores did not differ significantly (DISCERN, p = 0.188; EQIP, p = 0.935).

LLMs show promise in supporting patient education on AMI. While ChatGPT-4o offers superior accuracy and reliability, Claude 3.7 enhances accessibility through clearer language. This is the first study comparing three LLMs for AMI education in an emergency context, underscoring that physician oversight remains essential for educational applications in emergency medicine.