Interpreting ECG Images with Multimodal Large Language Models

Cardiac diseases are one of the leading causes of death worldwide. Electrocardiography (ECG) is one of the major diagnostic methods to detect cardiac diseases, including historical and acute myocardial infarction and arrhythmias. Recent advancements in AI, especially the introduction of multimodal large language models capable of understanding images, may provide a high level of accuracy in interpreting ECG images for possible cardiac diseases. We investigated Google Gemini 2.5 Pro and OpenAI GPT-4o for their abilities in recognizing the cardiac conditions on a corpus of 928 annotated ECG images and experimented on 10-shot and 5-shot classification methods, both with labelled examples to the models. In the 5-shot approach, examples fed to the LLMs agree with the annotation from a medical resident, and for 10-shot method, examples were randomly chosen from the train set. Our results show that few-shot classification methods significantly improved the accuracy with Gemini 2.5 Pro achieving 0.635 (10-shot) and 0.543 (5-shot) in accuracies, and GPT-4o reaching 0.819 (10-shot) and 0.829 (5-shot) accuracy, demonstrating their ability to interpret ECG images. Multimodal large language models may provide a cost effective and accurate way to interpret ECG images.