CADENCE Interprets ECG AI Models with Cardiac Atom Dictionary

Foundation models trained on 12-lead electrocardiograms (ECGs) have shown strong performance across various clinical tasks, but their internal workings and the physiological insights they capture often remain obscure. A new framework, named CADENCE, aims to demystify these models by breaking down their complex representations into a human-interpretable, queryable dictionary of "cardiac atoms." CADENCE employs a sparse autoencoder to factorize deep embeddings from millions of ECG tokens into thousands of these discrete cardiac atoms. These atoms have been shown to align more effectively with clinical phenotypes and waveform morphology than individual dense embedding dimensions. This allows for the recovery of specific medical conditions like arrhythmias, conduction abnormalities, and infarction patterns. The framework not only enhances the interpretability of ECG AI but also improves diagnostic accuracy, with atoms achieving higher AUROCs for clinical phenotypes and morphology compared to dense dimensions. CADENCE also enables targeted manipulation of model outputs and maintains consistent performance on external datasets, offering a scalable method for auditing and understanding AI in cardiology.