New framework decodes ECG foundation models into interpretable cardiac atoms

Researchers have developed CADENCE, a framework designed to make foundation models for electrocardiograms (ECGs) more interpretable. By using a BatchTopK sparse autoencoder, CADENCE breaks down ECG model embeddings into 8,192 distinct "cardiac atoms." These atoms better represent clinical phenotypes and waveform morphology compared to standard dense embeddings, achieving higher accuracy in identifying arrhythmias, conduction abnormalities, and other cardiac patterns. The framework also includes an LLM pipeline for generating and validating descriptions of these atoms, offering a scalable method for auditing the physiological knowledge within ECG foundation models. AI

IMPACT Enhances interpretability of AI models in medical diagnostics, potentially improving trust and auditing of clinical AI systems.

RANK_REASON The cluster describes a new research paper detailing a novel framework for interpreting AI models in a specific domain (ECGs). [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• BatchTopK
• CADENCE
• cardiac atoms
• foundation models
• LLM

• paper
• model release
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →