Few-shot ECG analysis with large vision language models in data-scarce clinical settings

Electrocardiogram (ECG) interpretation is fundamental for cardiac diagnosis. Machine learning has proven strong performance in ECG analysis but models often face challenges when deployed in real-world settings due to difficulties adapting to new environments. Recent progress, particularly through training on large, diverse datasets across various ECG-related tasks, has led to better model generalization. However, many cardiac conditions, particularly rare ones, remain scarce in existing datasets, leading to diminished model performance in these critical cases. This underrepresentation remains a major hurdle, as acquiring high-quality annotations for rare conditions is time-consuming and demands substantial clinical expertise and resources. In this paper, we investigate whether this issue can be addressed with the help of Large Vision-Language Models (VLMs) and In-Context Learning (ICL). We propose leveraging the inherent capabilities of VLMs, aided with only a few task-specific examples in the prompt to interpret ECG images in data-scarce environments, thus reducing reliance on extensive labeled data for fine-tuning. We introduce a real-world use case involving Brugada syndrome and evaluate the performance of pre-trained VLMs, comparing them to state-of-the-art ECG machine learning models. Results show that VLMs achieve competitive accuracy and, in data-constrained scenarios, outperform existing methods without requiring any updates to model weights. We further analyze the role of prompt engineering and input representation in influencing model performance. Our findings suggest that VLMs could serve as an alternative to address rare cardiac conditions, which are frequently overlooked because of data scarcity, positioning them as valuable assets for screening and triage-oriented analysis.