Interpretable multimodal zero shot ECG diagnosis via structured clinical knowledge alignment.

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Medical Imaging and Diagnostics

Background:

• Automated electrocardiogram (ECG) interpretation is crucial for cardiovascular disease diagnosis.
• Current AI systems lack transparency and generalization capabilities for novel conditions.
• Clinical workflows require interpretable diagnostic tools.

Purpose of the Study:

• To introduce ZETA, a zero-shot multimodal framework for interpretable ECG diagnosis.
• To align AI-driven ECG analysis with clinical differential diagnosis workflows.
• To enhance transparency, generalization, and trustworthiness of AI diagnostic systems.

Main Methods:

• Developed ZETA, a zero-shot multimodal framework utilizing a pre-trained model.
• Aligned ECG signals with structured positive/negative clinical observations via LLM-assisted curation.
• Mimicked differential diagnosis by comparing ECG embeddings with clinical text embeddings.
• Evaluated zero-shot classification performance and interpretability without disease-specific fine-tuning.

Main Results:

• ZETA achieved competitive zero-shot classification performance on ECG interpretation.
• Demonstrated enhanced interpretability, grounding predictions in clinically relevant features.
• Provided qualitative and quantitative evidence of improved diagnostic transparency.
• Showcased alignment of ECG analysis with structured clinical knowledge.

Conclusions:

• ZETA offers a transparent and generalizable approach to AI-powered ECG diagnosis.
• The framework mimics clinical differential diagnosis for improved interpretability.
• Aligning AI with structured clinical knowledge enhances trustworthiness in cardiovascular diagnostics.