🤖 AI

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

The paper introduces CADENCE, an interpretable framework that uses a sparse autoencoder to decompose ECG foundation model embeddings into a dictionary of 8,192 human-understandable "cardiac atoms" that outperform dense dimensions in predicting clinical phenotypes and morphology while providing transparent attribution for model decisions.

Original authors: Yixuan Duan, Arjun Naik, Sadeer Al-Kindi, Wei Qiu

Published 2026-07-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yixuan Duan, Arjun Naik, Sadeer Al-Kindi, Wei Qiu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a super-smart robot that can listen to your heart's electrical song and tell you if something is wrong. This robot is a "foundation model," a type of artificial intelligence trained on millions of heart recordings. It's incredibly good at spotting problems like irregular heartbeats or heart attacks, often better than a human doctor. But here's the catch: the robot is a "black box." It gives you an answer, but it keeps its reasoning locked inside a giant, tangled web of numbers. It's like a chef who makes a perfect cake but refuses to tell you the recipe, or even which ingredients they used. Doctors need to know why the robot made a decision to trust it, but right now, the robot's brain is just a blurry cloud of data.

To understand how this paper helps, we need to know two things. First, AI models often pack many different ideas into single "neurons," making it impossible to tell which idea is which. Second, scientists have recently found a way to untangle these neurons using a tool called a "sparse autoencoder." Think of this tool as a magical dictionary that takes a messy, mixed-up sentence and breaks it down into a list of clear, single words. This paper asks: Can we use this dictionary trick to open up the heart-robot's brain and find the specific "words" it uses to understand heart health?

The researchers created a new tool called CADENCE (Cardiac Atom Dictionary for Explainable Neural Concept Extraction) to solve this mystery. They took a powerful, pre-trained heart AI and fed its internal "thoughts" into a special decoder. Instead of getting a jumbled mess, the decoder spit out a dictionary of 8,192 "cardiac atoms." These aren't just random numbers; they are like tiny, specific building blocks of heart knowledge. Some atoms light up only when the heart beats too slowly, others only when a specific part of the heartbeat wave looks weird, and some only when the heart is in a specific rhythm.

The team found that these atoms are much better at spotting heart problems than the robot's original, messy brain cells. When they tested the atoms against real heart conditions, the best atoms were able to identify things like atrial fibrillation or heart block with an accuracy score (AUROC) of 0.88 to 0.90, whereas the original robot's raw data only scored around 0.78 to 0.83. It's as if they took a blurry photo and suddenly sharpened it, revealing details that were always there but hidden.

What makes CADENCE truly special is that it doesn't just find the problems; it explains them. The researchers used an automated AI assistant (a Large Language Model) to look at the heartbeats that made each atom "light up" and write a plain-English description of what it saw. For example, one atom was described as firing on the "elongated diastolic baseline between beats," which is the medical way of saying it spots a slow heart rate. Another atom was found to fire specifically on the "irregular fibrillatory baseline," identifying atrial fibrillation. The team proved these descriptions were real by having the AI guess which heartbeats would trigger the atom just by reading the description, and it got it right most of the time.

The paper also showed that these atoms behave like a logical map. If you look at how the atoms relate to each other, they naturally group together in ways that match human medical knowledge. For instance, atoms that detect a wide heartbeat wave (a sign of a conduction problem) sit right next to atoms that detect a long pause between beats, because in real heart physiology, those two things often happen together. Even better, when the researchers "turned off" a single atom, the robot's prediction for that specific heart problem dropped, proving that the atom was actually doing the work.

In short, CADENCE turns a mysterious, opaque heart AI into a transparent, searchable library. It shows that the robot isn't just guessing; it has learned a dictionary of real heart concepts, from the timing of the whole beat to the tiny shape of a single wave. By breaking the robot's brain down into these clear, named atoms, the researchers have given doctors a way to peek inside the black box, verify the robot's logic, and trust its advice with a much clearer understanding of the heart's secrets.

Technical Summary: CADENCE

Problem Statement

Electrocardiogram (ECG) foundation models (FMs) have demonstrated strong transferability across clinical tasks such as arrhythmia detection and risk prediction. However, these models operate as "black boxes," utilizing high-dimensional, dense embeddings where the internal physiological knowledge remains opaque. This lack of interpretability hinders clinical trust, auditing, and scientific utility. A specific challenge in interpreting these models is superposition: neural networks often encode more features than they have dimensions, causing individual neurons to be polysemantic (responding to multiple unrelated features) rather than mapping cleanly to single clinical concepts. While Sparse Autoencoders (SAEs) have successfully decomposed superposition in language, protein, and genomic foundation models, it remains an open question whether this approach can recover interpretable, monosemantic features from physiological time-series FMs and whether these features map to real cardiology.

Methodology

The authors present CADENCE (Cardiac Atom Dictionary for Explainable Neural Concept Extraction), a framework designed to decompose the latent representations of a frozen ECG foundation model into a human-interpretable dictionary of physiological concepts.

1. Backbone and Data

Model: The framework interprets a pre-trained 12-lead ECG foundation model (CSFM), a 6-layer transformer with a hidden width of 768.
Data: The dictionary is trained on the MIMIC-IV-ECG dataset (150,000 subject-disjoint ECGs for training, ~15,000 for testing). External validation is performed on the independent PTB-XL cohort.
Input: The model processes 12-lead ECGs tokenized into 100 ms patches, generating 1,200 tokens per recording. The residual-stream activations at all 6 layers are extracted as targets for dictionary learning.

2. Sparse Autoencoder (SAE) Architecture

Algorithm: CADENCE employs a BatchTopK Sparse Autoencoder. Unlike standard SAEs, BatchTopK retains the $k$ largest activations per token across the batch, fixing the average activation budget while allowing the number of active atoms to vary per token.
Configuration: The dictionary size ( $F$ ) is set to 8,192 atoms, with an average active code count ( $k$ ) of 128. This configuration was selected via grid search to balance reconstruction explained variance, the fraction of "dead" atoms, and interpretability yield.
Training: The SAE is trained on approximately 9 million tokens (Layer-6 embeddings) to reconstruct the input embeddings. The objective minimizes reconstruction error with an auxiliary term to mitigate dead atoms.

3. Concept Discovery and Validation Pipeline

Atom-Concept Association: Atoms are associated with clinical concepts (phenotypes, morphology, demographics) using rank-based AUROC scores on per-ECG maximum activations.
Relationship Analysis: Concept relationships are inferred from the cosine similarity of "prototype fingerprints" (mean $z$ -scored atom activations) without relational supervision.
Downstream Prediction: Linear probes (elastic-net logistic regression) are fitted on sparse atom activations and compared against probes on the original dense embeddings.
Automated Annotation: A Large Language Model (LLM) pipeline describes each atom based on its top-activating ECG crops and activation geometry (lead and beat-phase histograms). The LLM's descriptions are quantitatively validated by having the model predict held-out atom activations solely from the generated text.
Ablation Studies: Targeted ablation of specific atoms is performed to observe selective changes in frozen downstream model outputs, testing the functional necessity of individual atoms.

Key Contributions

First Application to Physiological Time-Series: To the authors' knowledge, this is the first application of SAE dictionary learning to a physiological time-series foundation model, demonstrating that ECG embeddings are monosemantically decomposable.
The CADENCE Framework: Introduces a method to factorize ECG foundation model representations into ~8,000 sparse cardiac atoms that are far more interpretable than raw neurons while faithfully reconstructing the representation.
Performance vs. Interpretability: Demonstrates that sparse atom probes match or exceed the predictive accuracy of dense embedding probes for phenotype, morphology, and age prediction, attributing predictions to small sets of named atoms.
Scalable Automated Interpretation: Establishes an automated LLM pipeline that generates and quantitatively validates human-readable descriptions for thousands of atoms, extending interpretability from hand-read exemplars to the full dictionary.
External Validity: Shows that the frozen MIMIC-trained dictionary transfers to the PTB-XL cohort without retraining, recovering overlapping physiological concepts and maintaining prediction performance.

Results

Concept Decodability: Clinical concepts become more linearly accessible with network depth. At Layer 6, the best SAE atoms achieve mean AUROCs of 0.88 for clinical phenotypes and 0.90 for waveform morphology, significantly outperforming the best dense embedding dimensions (0.78 and 0.83, respectively).
Predictive Accuracy: Sparse atom probes improve test AUROC for phenotype prediction from 0.93 (dense) to 0.95 (sparse) and maintain parity or improvement for morphology and age.
Hierarchical Organization: The recovered atoms span three levels of abstraction:
- Rhythm level: Multi-beat timing (e.g., sinus bradycardia firing on prolonged TP/RR intervals).
- Phenotype level: Discrete pathological events (e.g., PVCs firing on single premature beats; Atrial Fibrillation firing on fibrillatory baselines).
- Morphology level: Sub-beat waveform primitives (e.g., ST-segment deviations, PR prolongation).
Geometric Coherence: Atom-space geometry recovers physiologically coherent relationships (e.g., grouping LBBB, RBBB, and wide-QRS concepts) without explicit relational supervision.
Causal Control: Ablating a single target atom (e.g., an LBBB atom) selectively reduces the corresponding downstream logit, while ablating random or orthogonal atoms has no effect.
LLM Validation: The automated LLM pipeline achieves a median Spearman correlation ( $\rho$ ) of 0.65 between predicted and measured activations for held-out atoms, confirming the quantitative faithfulness of the generated descriptions.

Significance and Claims

The paper claims that CADENCE establishes a scalable framework for discovering and auditing the physiological knowledge encoded by ECG foundation models. By transforming an opaque model into a transparent, queryable dictionary of physiological concepts, CADENCE provides a foundation for:

Scientific Analysis: Surfacing hidden physiological structures and generating hypotheses about how models represent cardiac physiology.
Trustworthy Clinical AI: Enabling practitioners to inspect the specific waveform evidence (atoms) underlying a prediction, decompose diagnoses into named features, and identify confounded or shortcut features.
Generalizability: Demonstrating that the learned dictionary captures properties of the representation that generalize across datasets (MIMIC to PTB-XL), suggesting these are fundamental features of the ECG foundation model rather than dataset artifacts.

The authors conclude that sparse dictionary learning offers a practical route to auditing what physiological foundation models encode, moving beyond "black box" predictions toward transparent, scientifically grounded clinical AI.