CADENCE: A Cardiac Atom Dictionary for Interpretable Neural Concept Extraction from ECG Foundation Models

As ECG foundation models (FMs) become increasingly central to clinical AI, their internal representations remain opaque, often described as "black boxes." This study asks whether the high-dimensional, polysemantic embeddings of these models can be decomposed into a human-interpretable dictionary of physiological concepts, and whether such a dictionary can maintain or improve predictive performance while providing clinical auditability.

The authors introduce CADENCE (Cardiac Atom Dictionary for Explainable Neural Concept Extraction). They apply a BatchTopK sparse autoencoder (SAE) to the Layer-6 residual stream of a pre-trained 12-lead ECG foundation model. By factorizing these embeddings into 8,192 sparse "cardiac atoms," the researchers create a queryable dictionary. They evaluate the framework by measuring how well these atoms align with known clinical phenotypes (e.g., arrhythmias, conduction abnormalities) and waveform morphology, and by using an automated LLM pipeline to generate and validate human-readable descriptions for each atom based on its activation patterns.

CADENCE successfully disentangles the model's internal representations, revealing that sparse atoms align significantly better with clinical concepts than the original dense embedding dimensions. At Layer 6, the best atoms achieve a mean AUROC of 0.88 for clinical phenotypes and 0.90 for morphology, compared to 0.78 and 0.83 for dense dimensions. Furthermore, sparse atom-based probes match or outperform dense probes in downstream tasks, improving phenotype prediction AUROC from 0.93 to 0.95. The framework allows for targeted ablation, where removing specific atoms selectively alters model outputs, and it demonstrates consistent performance across independent external datasets.

This work provides a scalable, automated method for auditing clinical AI, transforming opaque neural representations into a transparent, scientifically grounded dictionary. By enabling practitioners to inspect the specific physiological features driving a model's prediction, CADENCE bridges the gap between high-performance deep learning and the clinical requirement for explainable, trustworthy diagnostic tools.