Abstract
BACKGROUND Vision-enabled large language models (VE-LLMs) have the potential to provide flexible and explainable medical image interpretation. However, their real-world performance on clinical data such as 12-lead electrocardiograms (ECGs) has not been systematically assessed. OBJECTIVE This study aimed to evaluate the diagnostic accuracy and reliability of state-of-the-art VE-LLMs in interpreting real-world ECG images. METHODS We tested eight VE-LLMs (ChatGPT-5, ChatGPT-4, Gemini 2.5 Pro, Copilot, Grok-4, Perplexity, Claude Sonnet-4, and Claude Opus-4.1) using 70 de-identified ECG images. A standardized prompt requested nine determinations: rhythm, first-degree atrioventricular (AV) block, intraventricular conduction block and pattern, corrected QT (QTc) prolongation, premature atrial and ventricular contractions, ischemic ST-segment deviation, and axis deviation. An expert consensus served as the reference standard. Model outputs were evaluated using overall and per-class diagnostic metrics. RESULTS Overall accuracy across models varied significantly from 68.1% to 78.3% (429/630 to 493/630, Cochran’s Q, P<.001). ChatGPT-5 achieved the highest accuracy (78.3%) but had the slowest response time (median 276 seconds), whereas Perplexity and Copilot responded within a median of 2 and 3 seconds, respectively. Rhythm classification reached 72.9%–82.9% accuracy (51/70 to 58/70), but sensitivity for atrial fibrillation was ≤22% (≤2/9). Detection of first-degree AV block was poor (sensitivity 0%–33%; 0/9 to 3/9), and QTc prolongation was also poor (sensitivity 0%–45.5%; 0/22 to 10/22). Intraventricular block was identified with up to 70% accuracy (49/70), but correct subtype assignment was ≤44% (≤11/25). ST-segment deviation sensitivity was <25% for all models (highest 3/14). Agreement with expert interpretation was low, with Cohen’s kappa (κ) indicating poor-to-fair concordance (κ≤.37). CONCLUSIONS VE-LLMs showed moderate overall accuracy but low sensitivity and limited agreement with expert ECG interpretation. Current performance is inconsistent and insufficient for clinical deployment. Future development should focus on domain-specific training and hybrid approaches combining LLM reasoning with established ECG algorithms before use in patient care.