Multi-View Echocardiographic Embedding for Accessible AI Development

Area of Science:

• Artificial Intelligence in Medicine
• Cardiovascular Imaging
• Machine Learning for Healthcare

Background:

• Echocardiography is crucial for cardiovascular diagnostics.
• Current AI foundation models for cardiac imaging are computationally intensive and data-hungry, limiting accessibility.
• Vector embeddings offer a solution for compact data representation in AI applications.

Purpose of the Study:

• To develop a computationally accessible multi-view encoder framework for cardiac AI.
• To investigate demographic fairness challenges in AI models for echocardiography.
• To improve diagnostic performance and efficiency in cardiovascular AI applications.

Main Methods:

• Developed a transformer-based multi-view encoder using the MIMIC-IV-ECHO dataset (7,169 studies).
• Aggregated view-level representations into study-level embeddings for efficient downstream tasks.
• Employed adversarial learning to mitigate demographic bias while preserving clinical performance across 21 classification tasks.

Main Results:

• The multi-view encoder achieved a mean improvement of 9.0 AUC points (12.0% relative improvement) compared to foundation model baselines.
• Performance remained robust even with a reduced number of echocardiographic views.
• Adversarial learning demonstrated limited success in eliminating demographic shortcuts without compromising diagnostic accuracy.

Conclusions:

• The developed framework democratizes advanced cardiac AI, offering substantial diagnostic improvements with reduced computational needs.
• The multi-view encoder provides a practical pathway for wider AI adoption in cardiovascular medicine.
• Enhanced efficiency and accessibility are key benefits for real-world clinical settings.