Answers are grounded in this paper and cite their sources.
BACKGROUND: Although recent artificial intelligence (AI) foundation models have shown superior performance in medical imaging, their high computational demands and reliance on large data sets limit accessibility. These issues are especially pronounced in echocardiography, where integrating complex multiview measurements for individual cases hinders AI development and clinical translation. Vector embedding approaches offer promising solutions by leveraging compact representations in AI applications. OBJECTIVES: We aimed to develop a transformer-based multiview encoder that integrates multiple echocardiographic videos into unified vector embeddings and uses adversarial learning to reduce reliance on demographic shortcuts. METHODS: Using the MIMIC-IV-ECHO (Medical Information Mart for Intensive Care IV-Echocardiography) data set (7,169 studies), we trained a multiview encoder to produce study-level embeddings and evaluated them on 21 binary classification tasks covering echocardiographic measurements and clinical diagnoses. We varied adversarial weights to assess their impact on demographic shortcut reliance and overall performance. RESULTS: The multiview encoder raised the mean area under the curve (AUC) across clinical tasks from 0.742 to 0.831 compared with the foundation model embedding (ΔAUC +0.090; 95% CI: 0.078-0.100), with significant gains in 20 of 21 tasks. The embeddings remained robust with a limited number of views (mean ΔAUC +0.071; 95% CI: 0.043-0.113). Adversarial learning reduced sex and race predictability only at strong weights and at substantial cost to clinical performance. CONCLUSIONS: The multiview encoder framework integrates multiple echocardiographic views into unified study-level embeddings, enabling substantial diagnostic improvements across diverse cardiac tasks. By producing compact, informative embeddings without requiring extensive computational resources, this approach lowers the cost of building echocardiographic AI on an existing foundation model.
No takes yet. Share an insight, caveat, or question.
Tohyama et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: