CADENCE Decomposes ECG Foundation Models into 8,192 Interpretable Cardiac Atoms
July 29, 2026
CADENCE uses a BatchTopK sparse autoencoder to factorize Layer-6 embeddings from nine million ECG tokens into 8,192 sparse atoms. These atoms achieve 0.88 AUROC for clinical phenotypes and 0.90 for morphology, outperforming dense embedding dimensions in representing arrhythmias and conduction abnormalities.
HOW THIS AFFECTS YOU
●
researcherYou can use sparse autoencoders to extract more interpretable physiological features from foundation model embeddings.
●
healthThis improves the transparency of AI-driven ECG diagnostics by mapping neural activations to clinical phenotypes.