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

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Machine Learning for Healthcare

Background:

• Electrocardiogram (ECG) interpretation is crucial for cardiac diagnosis.
• Machine learning models struggle with real-world ECG data generalization, especially for rare conditions due to data scarcity.
• Acquiring expert annotations for rare cardiac conditions is resource-intensive.

Purpose of the Study:

• To investigate the use of Large Vision-Language Models (VLMs) with In-Context Learning (ICL) for ECG interpretation in data-scarce environments.
• To reduce the reliance on extensive labeled datasets for training ECG analysis models.
• To evaluate VLM performance against state-of-the-art ECG machine learning models for rare cardiac conditions.

Main Methods:

• Leveraging pre-trained VLMs with task-specific examples provided in the prompt (ICL).
• Utilizing a real-world use case focusing on Brugada syndrome.
• Comparing VLM performance to existing state-of-the-art ECG machine learning models.

Main Results:

• VLMs achieved competitive accuracy in ECG interpretation.
• In data-constrained scenarios, VLMs outperformed existing methods without requiring model weight updates.
• Prompt engineering and input representation were analyzed for their impact on VLM performance.

Conclusions:

• VLMs offer a viable alternative for addressing rare cardiac conditions often overlooked due to data scarcity.
• VLMs can serve as valuable tools for screening and triage-oriented ECG analysis.
• This approach reduces the need for extensive fine-tuning on limited datasets.