Pith. sign in

REVIEW 5 major objections 6 minor 53 references

The hidden representations of a 12-lead ECG foundation model can be decomposed into a dictionary of 8,192 sparse, interpretable 'atoms' that align with clinical heart concepts better than the model's own neurons, without sacrificing predict

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 02:58 UTC pith:3DOMVDMB

load-bearing objection Solid first application of SAEs to an ECG foundation model; the headline atoms-vs-neurons gap is inflated by test-set selection, but the framework and qualitative claims survive. the 5 major comments →

arxiv 2607.25244 v1 pith:3DOMVDMB submitted 2026-07-28 cs.AI

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

Yixuan Duan , Arjun Naik , Sadeer Al-Kindi , Wei Qiu This is my paper
classification cs.AI
keywords sparse autoencodersmechanistic interpretabilityECG foundation modelsconcept discoveryelectrocardiogramrepresentation learninginterpretable AI in healthcare
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that the opaque embeddings of a 12-lead ECG foundation model are not irreducibly tangled: a BatchTopK sparse autoencoder can factor them into a dictionary of 8,192 sparse 'atoms,' each firing on a specific, localizable waveform feature. The atoms align substantially better than individual dense embedding dimensions with clinical phenotypes and waveform morphology (mean best-atom AUROC 0.88 and 0.90 at the deepest layer, vs. 0.78 and 0.83 for the best dense dimensions), and linear probes trained on atoms match or outperform probes on the dense embedding while attributing every prediction to a small set of named atoms (phenotype AUROC improves from 0.93 to 0.95). The same atoms and their physiological relationships reappear on an independent external ECG dataset without retraining, and single-atom ablation selectively moves frozen downstream predictions. If the claim holds, ECG foundation models become auditable and queryable — practitioners can see which waveform evidence a diagnosis rests on and even use the dictionary to surface new, sub-diagnostic morphology.

Core claim

The central discovery is that a sparse autoencoder trained on the residual-stream activations of a frozen ECG foundation model decomposes the representation into a structured dictionary of cardiac atoms that are far more monosemantic than raw neurons. At the deepest layer, the best single atom detects clinical phenotypes with a mean AUROC of 0.88 and waveform morphology with 0.90, compared with 0.78 and 0.83 for the best dense embedding dimension. The atoms recover arrhythmias, conduction abnormalities, infarction and repolarization patterns, chamber/axis findings, and lead- and beat-phase-specific waveform primitives; atom-space fingerprints self-organize into physiologically coherent block

What carries the argument

The central mechanism is the BatchTopK sparse autoencoder: it encodes each 768-dimensional token embedding from the model's deepest layer into a dictionary of 8,192 atoms with a fixed average active budget of k=128 codes, and reconstructs the embedding from those atoms. The decoder's unit-normalized columns are the 'atoms,' each of which can be overlaid on the raw ECG to show precisely what waveform feature it detects. This dictionary is then used as a substrate for four analyses: concept association via per-ECG max-activation AUROC fingerprints, sparse linear probing for downstream prediction with per-atom attributions, concept-relationship geometry via fingerprint cosines, and automated LL

Load-bearing premise

The claim that atoms align with clinical concepts better than raw neurons rests on comparing the single best atom out of 8,192 candidates with the single best dense unit out of 768, both selected after seeing the test set; without a correction for this selection, part of the reported gap may be an artifact of searching more candidates.

What would settle it

Re-run the best-atom versus best-dense comparison with the best atom selected on a training split (or with a multiple-comparison correction); if the AUROC gap between atoms and dense dimensions collapses toward zero, the central claim of superior concept alignment is falsified. A second check: train a sparse autoencoder on random permutations of the same embeddings and ask whether its atoms still align with clinical concepts — they should not, if the alignment reflects true physiological structure.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If ECG foundation-model embeddings are monosemantically decomposable, any such model can be turned into an auditable, queryable concept dictionary without retraining or fine-tuning.
  • Diagnosis predictions become attributable: each prediction decomposes into a small set of named atoms whose activation can be checked against the waveform by a clinician.
  • The recovery of known physiological relationships (conduction blocks clustering with wide QRS, atrial fibrillation with flutter) without relational supervision suggests the dictionary can generate new hypotheses about cardiac structure.
  • Because the frozen dictionary transfers to an independent external cohort, the recovered concepts appear to be properties of the representation itself, not dataset artifacts.
  • Single-atom ablation steering shows that atoms function as selective control knobs on frozen downstream logits, enabling causal-style interrogation of the model's decisions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The headline atoms-vs-dense advantage may be partly inflated by selection: the best of 8,192 atoms is compared with the best of 768 dense dimensions, both picked on the test set, without a multiple-testing correction. A held-out selection protocol would reveal how much of the gap is real.
  • If the dictionary truly captures the representation's units, the same method should extend to other physiological time-series foundation models (e.g., PPG, EEG, single-lead ECGs), where the recovered atoms might reveal cross-modal physiological primitives.
  • The LLM description-validation loop suggests a scalable audit mechanism: atom descriptions that pass activation-prediction could be used to scan for shortcut features or demographic confounds before deployment.
  • The finding that atoms increasingly outperform dense dimensions with depth implies superposition is more pronounced in shallow layers; attaching interpretability tools at the right depth may be as important as the dictionary itself.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper introduces CADENCE, a BatchTopK sparse autoencoder (SAE) trained on the frozen Layer-6 residual-stream embeddings of an ECG foundation model (CSFM) over roughly nine million patch tokens from MIMIC-IV-ECG. The learned 8,192 sparse atoms are evaluated as concept detectors: for each of 16 MIMIC-derived clinical phenotypes, morphology measurements, and age, the paper reports the best single-atom AUROC across layers and compares it with the best single dense CSFM dimension, claiming large advantages at Layer 6 (e.g., phenotype 0.88 vs. 0.78). Sparse linear probes on atom activations are reported to match or exceed dense embedding probes for phenotype, morphology, and age prediction, while providing per-atom attributions. Additional analyses include atom-space concept relationships, single-atom ablation steering of frozen downstream predictors, an LLM-based describe-then-validate annotation pipeline, and external validation on PTB-XL using the frozen MIMIC-trained SAE.

Significance. If the central quantitative claims survive scrutiny, this would be an important contribution: it would be the first SAE-based dictionary analysis of a physiological time-series foundation model, showing monosemantic decomposition into physiologically meaningful atoms, with substantial supporting evidence including open code, detailed appendices, a controlled ablation audit, an external cardiologist-labeled cohort, and a quantitative LLM validation protocol. The paper is not circular in the derivation sense: the SAE is trained unsupervised, and most predictive claims are evaluated on held-out subjects. However, several load-bearing comparisons currently rely on test-set selection without correction, and the main MIMIC phenotype labels are keyword-derived from machine-report text; these issues materially affect the strength of the central claims and require revision before the paper can be recommended for acceptance.

major comments (5)
  1. [§3.4, Table A1, App. A.2] The headline claim that atoms align substantially better than dense dimensions rests on selecting the single best sign-folded feature on the full test set from dictionaries of different sizes (8,192 atoms vs. 768 dense units, doubling to 16,384 vs. 1,536 after sign folding). This maximum-of-many vs. maximum-of-few comparison is not selection-corrected. The bootstrap CIs reported in Table A1 condition on the already-selected feature and therefore do not capture selection uncertainty. For weak-signal concepts such as LVH (Table A1, L6: SAE 0.684 vs. dense 0.631) or low QRS voltage (0.712 vs. 0.654), the gap is small and within plausible selection noise. The authors should either split the test set into selection and evaluation folds, choosing the best feature on the selection fold and reporting AUROC on the evaluation fold, or apply a permutation-based multiple-testing correction (e.g., We
  2. [App. A.1, App. A.10] The SAE hyperparameters (F=8192, k=128) are selected by a grid search scored in part on the 'number of phenotype-specific atoms' (App. A.1). The held-out evaluation set is used for all downstream analyses (App. A.10), and no separate validation split is described for this grid search. If the same held-out set is used both to select F and k and to report best-atom AUROC, the reported numbers are optimistic. Model selection should be performed on a validation fold, with the final evaluation on a disjoint test fold; at minimum, the paper should state explicitly which split the grid search used and, if it used the test set, re-run the evaluation with a proper split.
  3. [App. A.8, Table A4, §3.1] MIMIC phenotype labels are derived by keyword matching on machine-report text, not by independent cardiologist adjudication. The main 'clinical phenotype' AUROCs therefore measure association with report keywords, and the statement that atoms 'align with clinical phenotypes' overstates the ground truth. The external PTB-XL analysis uses cardiologist-adjudicated superclasses, but only for five diagnostic classes; the 49 measurement concepts are automated measurements. The authors should either (a) validate on an adjudicated subset of MIMIC or (b) clearly qualify in the abstract and results that the headline MIMIC phenotype numbers are relative to machine-report text, with external adjudicated results as the stricter check. This is load-bearing for the physiological-interpretability claim.
  4. [§3.5, Fig. 2b, App. A.3] The probe comparison uses max-pooled SAE atom activations (8192-d) against mean-pooled dense CSFM embeddings (768-d). Since many phenotypes are temporally localized, max-pooling can give the atom probe a systematic advantage unrelated to dictionary learning. The paper does not report dense probes with max-pooling, which would be the natural control. The claim that 'sparse probes match or exceed dense probes' should be supported by comparing both pooling choices for both representations, or by a clearly justified asymmetry. This is particularly relevant for the shallow-layer gains (e.g., phenotype 0.843 vs. 0.805 at Layer 1).
  5. [App. A.7, Fig. A15] The LLM describe-then-validate protocol is described as 'geometry-first' and free of report-label leakage, but the per-atom user prompt includes 'report enrichment = {...}, best phenotype assoc = <pheno> (AUROC <x>)' (Fig. A15). The ablation in App. A.7 only tests adding the full report text to the prompt; it does not test removing the already-included enrichment and phenotype-association fields. The validation may therefore reflect information from the same report-derived labels used elsewhere in the paper. To support the claim of quantitative, label-independent description validation, the prompt should be rerun without these auxiliary fields, or the leakage analysis should directly compare with and without them.
minor comments (6)
  1. [Throughout] The paper alternates between 'neurons' and 'dense dimensions'/'dense units' when referring to the raw CSFM embedding coordinates. Since the CSFM is a transformer, the 768-d embedding dimensions are not neurons in the usual sense; please use 'dense units' consistently.
  2. [Abstract and App. A.10] The abstract says 'more than nine million tokens', while App. A.10 gives 150,000 ECGs × 60 tokens = 9,000,000. The wording 'more than' is slightly inaccurate unless additional tokens are used; please reconcile.
  3. [Sec. 4.1 and App. A.2] The main text says 'we defer this atoms-vs-neurons comparison to Appendix A.2' but already reports the comparison numbers in Sec. 4.1. Please avoid the deferred-comparison phrasing or move the detailed numbers to the appendix.
  4. [App. A.7] The LLM name 'gpt-5.4-mini' is used without a citation or version identifier. Since this model is part of the method, please provide a reference or clear model card details.
  5. [App. A.10] The subject-disjoint split uses hash(subject_id) mod 104 < 1500. The choice of 104 and the modulo threshold should be justified, as hashing may not be perfectly uniform across the actual subject-id distribution.
  6. [Fig. 5a] The distribution of Spearman rho is shown for 1000 atoms; please state explicitly that the 92% above 0.3 and 2% below 0 are computed on this sample, and include the sample size in the caption.

Circularity Check

1 steps flagged

Best-atom vs. best-dense comparison is selected on the full test set, partially forcing the headline atom advantage; the remaining derivation is independent.

specific steps
  1. fitted input called prediction [Sec. 3.4; Table A1 caption, App. A.3]
    "The best-atom score of a concept at a layer is the largest such association over the dictionary, evaluated on the held-out test set. ... For each concept and layer we select, on the full test set, the single best sign-folded SAE atom and the single best sign-folded CSFM dense unit, then bootstrap that fixed unit's AUROC by resampling patients."

    The reported 'best-atom AUROC' is by definition the maximum over 8,192 per-atom AUROCs computed on the same test set used to choose the atom, while the dense baseline is the maximum over only 768 units. The reported value is thus an in-sample, test-set-optimized maximum, not a held-out prediction of the selected atom; the 0.88 vs 0.78 gap is partly forced by the ~10.7x larger candidate pool (max-of-many vs max-of-few). The paper discloses the selection, and probes/ablation/PTB-XL are independent, so the circularity is partial and evaluation-level.

full rationale

The core derivation is not circular: the BatchTopK SAE is trained unsupervised on frozen CSFM Layer-6 tokens (Eq. 1-2), atom-concept associations are computed post hoc, sparse probes are fit and evaluated on subject-disjoint splits, ablation steering uses a frozen downstream predictor, and the LLM validation predicts held-out activations from descriptions. No fitted parameter is renamed as a prediction in these parts, and no load-bearing self-citation appears. The one genuine circular/selection artifact is the best-atom-versus-best-dense comparison: the paper selects the single best atom and the single best dense unit on the full test set (Table A1 caption) and then reports that fixed unit's test AUROC, so the headline 'atoms align better than neurons' is an in-sample maximum over 8,192 candidates against 768, inflated by candidate count. Hyperparameter choice also uses a label-derived 'phenotype-specific atom count' proxy, but the paper does not explicitly state that this grid search uses the test set, so I do not count it as a separate circular step. Because the SAE representation, probe parity, geometry, steering, and external transfer are independently evaluated, the overall circularity is partial and centered on the headline atom-vs-dense claim; score 5.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 1 invented entities

The central claim rests on the SAE assumption, noisy keyword-derived labels, a Layer-6 choice made after inspecting depth trends, and an evaluation procedure that selects the best of 8,192 atoms on the test set. These are not fatal, but they mean the headline numbers are partly optimized rather than purely measured.

free parameters (3)
  • Dictionary size F = 8192
    Chosen by grid search over F∈{4096,8192,16384,32768} and k∈{32,64,128,256}; selection criterion included number of phenotype-specific atoms, so hyperparameters are partially tuned to maximize the paper's headline interpretability result (App. A.1, Fig. A2).
  • Active code budget k = 128
    Chosen in the same grid search; k=128 maximizes phenotype-specific atom count at F=8192, the interpretability proxy used to select hyperparameters (App. A.1).
  • LLM atom inclusion thresholds = dead/rare (<20 high-activation ECGs)/near-constant (>90% ECGs)/weak/lead-diffuse (entropy>0.85)
    Pre-hoc filters determine which 2991 of 8192 atoms are interpreted and validated; thresholds are motivated by a pilot but not independently justified (App. A.7).
axioms (6)
  • domain assumption BatchTopK sparse autoencoders recover monosemantic features from polysemantic neural activations.
    Cadence's central mechanism; inherited from SAE literature (Refs 5,8,45), not proven in this paper.
  • domain assumption MIMIC-IV-ECG machine-report keyword labels are adequate proxies for clinical phenotypes.
    All phenotype associations and probe labels use keyword-matched machine-report text (Table A4), not cardiologist-adjudicated labels for MIMIC.
  • domain assumption Layer-6 residual stream is the right representation to factorize.
    Layer 6 is chosen because concepts are most decodable there, but this choice is post hoc after seeing depth trends (Sec. 4.1).
  • ad hoc to paper Selecting the best atom per concept on the test set and comparing across dictionaries of different sizes yields valid concept-alignment estimates.
    Table A1 selects best units on the full test set; no multiple-testing correction; central comparison is between 8192 and 768 candidates.
  • domain assumption LLM predictions of held-out activations from text are a valid measure of description faithfulness.
    Spearman correlation between LLM ordinal scores and activations; assumes the model is not exploiting generic ECG priors (dissimilar-atom control provided).
  • domain assumption External PTB-XL labels and measurements are independent and accurate.
    PTB-XL cardiologist-adjudicated diagnoses and machine measurements are used as external validation; dataset labels are assumed trustworthy.
invented entities (1)
  • Cardiac atoms (8,192 sparse SAE features) no independent evidence
    purpose: Interpretable units supposed to correspond to physiological concepts such as PVC, AF, ST segment, and bundle branch blocks.
    Atoms are learned dictionary components of the model's representation, not independently observable entities; external transfer and ablation provide indirect evidence, but no falsifiable prediction outside the model is offered.

pith-pipeline@v1.3.0-alltime-deepseek · 25280 in / 11148 out tokens · 108388 ms · 2026-08-01T02:58:40.331352+00:00 · methodology

0 comments
read the original abstract

Foundation models for 12-lead electrocardiograms (ECGs) transfer well across clinical tasks, but the physiological knowledge encoded in their representations remains opaque. We present CADENCE, a framework that decomposes an ECG foundation model into a human-interpretable, queryable dictionary of physiological concepts. Using a BatchTopK sparse autoencoder, CADENCE factorizes Layer-6 embeddings from more than nine million ECG tokens into 8,192 sparse cardiac atoms. These atoms align better than individual dense embedding dimensions with clinical phenotypes and waveform morphology, recovering arrhythmias, conduction abnormalities, infarction and repolarization patterns, chamber and axis findings, and lead- and beat-phase-specific waveform primitives. At Layer 6, the best atoms achieve mean AUROCs of 0.88 for clinical phenotypes and 0.90 for morphology, versus 0.78 and 0.83 for the best dense dimensions. Sparse atom probes match or outperform dense probes for phenotype, morphology, and age prediction while attributing each prediction to a small set of interpretable atoms; phenotype AUROC improves from 0.93 to 0.95. Atom-space geometry recovers physiologically coherent relationships, and targeted atom ablation selectively changes frozen downstream outputs. An automated LLM pipeline generates and quantitatively validates atom descriptions by predicting held-out activations. On independent external ECG datasets, CADENCE recovers overlapping concepts and maintains consistent phenotype-prediction performance. CADENCE provides a scalable framework for discovering and auditing the physiological knowledge encoded by ECG foundation models.

Figures

Figures reproduced from arXiv: 2607.25244 by Arjun Naik, Sadeer Al-Kindi, Wei Qiu, Yixuan Duan.

Figure 1
Figure 1. Figure 1: Overview of CADENCE. (a) Dictionary construction. A frozen ECG foundation model encodes a 12-lead ECG into [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: What the dictionary encodes, how well it predicts, and where it fires. (a,b) Phenotype, morphology, and age across [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Phenotype relationships in atom space. Concept– [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Concept-level steering and sparse probing. (a) For four targets—LBBB, wide QRS, age, one-year mortality—the change [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Automated LLM atom descriptions are quantitatively faithful. (a) Per-atom Spearman [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

53 extracted references · 5 canonical work pages

  1. [1]

    Etowah Adams, Liam Bai, Minji Lee, Yiyang Yu, and Mohammed AlQuraishi

  2. [2]

    Guillaume Alain and Yoshua Bengio. 2016. Understanding Intermediate Layers Using Linear Classifier Probes.arXiv preprint arXiv:1610.01644(2016). https: //arxiv.org/abs/1610.01644

    Pith/arXiv arXiv 2016

  3. [3]

    Attia, Suraj Kapa, Francisco Lopez-Jimenez, Paul M

    Zachi I. Attia, Suraj Kapa, Francisco Lopez-Jimenez, Paul M. McKie, Dorothy J. Ladewig, Gaurav Satam, Patricia A. Pellikka, Maurice Enriquez-Sarano, Peter A. Noseworthy, Thomas M. Munger, Samuel J. Asirvatham, Christopher G. Scott, Rickey E. Carter, and Paul A. Friedman. 2019. Screening for Cardiac Contractile Dysfunction Using an Artificial Intelligence-...

    doi:10.1038/s41591-018-0240-2 2019

  4. [4]

    Steven Bills, Nick Cammarata, Dan Mossing, Henk Tillman, Leo Gao, Gabriel Goh, Ilya Sutskever, Jan Leike, Jeff Wu, and William Saunders. 2023. Language Models Can Explain Neurons in Language Models.OpenAI(2023). https: //openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html

    2023

  5. [5]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, et al. 2023. Towards Monosemanticity: Decomposing Language Models With Dictionary Learning.Transformer Circuits Thread(2023). https://transformer-circuits.pub/ 2023/monosemantic-features/index.html

    2023

  6. [6]

    Bart Bussmann, Patrick Leask, and Neel Nanda. 2024. BatchTopK Sparse Autoen- coders.arXiv preprint arXiv:2412.06410(2024). https://arxiv.org/abs/2412.06410

    Pith/arXiv arXiv 2024

  7. [7]

    Cook, So-Young Oh, and Martin V

    David A. Cook, So-Young Oh, and Martin V. Pusic. 2020. Accuracy of Physi- cians’ Electrocardiogram Interpretations: A Systematic Review and Meta-analysis. JAMA Internal Medicine180, 11 (2020), 1461–1471. doi:10.1001/jamainternmed. 2020.3989

    arXiv 2020

  8. [8]

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey

  9. [9]

    Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, et al. 2022. Toy Models of Superposition.Transformer Circuits Thread (2022). https://transformer-circuits.pub/2022/toy_model/index.html

    2022

  10. [10]

    Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2024. Scaling and Evaluating Sparse Autoencoders.arXiv preprint arXiv:2406.04093(2024). https://arxiv.org/ abs/2406.04093

    Pith/arXiv arXiv 2024

  11. [11]

    Goldberger, Luis A

    Ary L. Goldberger, Luis A. N. Amaral, Leon Glass, Jeffrey M. Hausdorff, Plamen Ch. Ivanov, Roger G. Mark, Joseph E. Mietus, George B. Moody, Chung-Kang Peng, and H. Eugene Stanley. 2000. PhysioBank, PhysioToolkit, and PhysioNet: Compo- nents of a New Research Resource for Complex Physiologic Signals.Circulation 101, 23 (2000), e215–e220. doi:10.1161/01.CI...

    doi:10.1161/01.cir.101.23.e215 2000

  12. [12]

    Nathanson, Alistair Johnson, Benjamin Moody, Cheryl Fernandes, Nathaniel Greenbaum, Jonathan W

    Brian Gow, Tom Pollard, Larry A. Nathanson, Alistair Johnson, Benjamin Moody, Cheryl Fernandes, Nathaniel Greenbaum, Jonathan W. Waks, Parastou Eslami, Tanner Carbonati, Ashish Chaudhari, Elizabeth Herbst, Dana Moukheiber, Seth Berkowitz, Roger Mark, and Steven Horng. 2023. MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset (version 1.0).PhysioNet(...

    2023

  13. [13]

    Gowda, Antonio H

    Xiao Gu, Wei Tang, Jinpei Han, Veer Sangha, Fenglin Liu, Shreyank N. Gowda, Antonio H. Ribeiro, Patrick Schwab, Kim Branson, Lei Clifton, Antonio Luiz P. Ribeiro, Zhangdaihong Liu, and David A. Clifton. 2026. Cardiac health assessment across scenarios and devices using a multimodal foundation model pretrained on data from 1.7 million individuals.Nature Ma...

    work page doi:10.1038/s42256-026-01180-5 2026

  14. [14]

    Haoxiang Guan, Jiyan He, and Jie Zhang. 2025. Sparse Autoencoders Re- veal Interpretable Structure in Small Gene Language Models.arXiv preprint arXiv:2507.07486(2025). https://arxiv.org/abs/2507.07486

    Pith/arXiv arXiv 2025

  15. [15]

    Wes Gurnee, Neel Nanda, Matthew Pauly, Katherine Harvey, Dmitrii Troitskii, and Dimitris Bertsimas. 2023. Finding Neurons in a Haystack: Case Studies with Sparse Probing.Transactions on Machine Learning Research (TMLR)(2023). https://arxiv.org/abs/2305.01610

    Pith/arXiv arXiv 2023

  16. [16]

    Hanley and Barbara J

    James A. Hanley and Barbara J. McNeil. 1982. The Meaning and Use of the Area Under a Receiver Operating Characteristic (ROC) Curve.Radiology143, 1 (1982), 29–36. doi:10.1148/radiology.143.1.7063747

    doi:10.1148/radiology.143.1.7063747 1982

  17. [17]

    Hannun, Pranav Rajpurkar, Masoumeh Haghpanahi, Geoffrey H

    Awni Y. Hannun, Pranav Rajpurkar, Masoumeh Haghpanahi, Geoffrey H. Tison, Codie Bourn, Mintu P. Turakhia, and Andrew Y. Ng. 2019. Cardiologist-Level Arrhythmia Detection and Classification in Ambulatory Electrocardiograms Using a Deep Neural Network.Nature Medicine25, 1 (2019), 65–69. doi:10.1038/ s41591-018-0268-3

    2019

  18. [18]

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick

  19. [19]

    Gerhard Hindricks, Tatjana Potpara, Nikolaos Dagres, et al . 2021. 2020 ESC Guidelines for the diagnosis and management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS).European Heart Journal42, 5 (2021), 373–498. doi:10.1093/eurheartj/ ehaa612

    doi:10.1093/eurheartj/ 2021

  20. [20]

    Alistair E. W. Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J. Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, Li-wei H. Lehman, Leo A. Celi, and Roger G. Mark. 2023. MIMIC-IV, a freely accessible electronic health record dataset.Scientific Data10, 1 (2023), 1. doi:10.1038/s41597- 022-01899-x

    doi:10.1038/s41597- 2023

  21. [21]

    Ihor Kendiukhov. 2026. Sparse autoencoders reveal organized biological knowl- edge but minimal regulatory logic in single-cell foundation models: a compar- ative atlas of Geneformer and scGPT.arXiv preprint arXiv:2603.02952(2026). https://arxiv.org/abs/2603.02952

    arXiv 2026

  22. [22]

    Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viégas, and Rory Sayres. 2018. Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV). InProceedings of the 35th International Conference on Machine Learning (ICML). 2668–2677. https://arxiv.org/abs/1711.11279

    Pith/arXiv arXiv 2018

  23. [23]

    Koo, and Su-In Lee

    Chanwoo Kim, Jakub Kaczmarzyk, Deepika Savant, Zhen Zhao, Peter K. Koo, and Su-In Lee. 2026. Dissecting and directing pathology foundation models.bioRxiv (2026). doi:10.64898/2026.06.12.731496

    work page doi:10.64898/2026.06.12.731496 2026

  24. [24]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochastic Optimization. InInternational Conference on Learning Representations (ICLR). https://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2015

  25. [25]

    Dani Kiyasseh, Tingting Zhu, and David A. Clifton. 2021. CLOCS: Contrastive Learning of Cardiac Signals Across Space, Time, and Patients. InProceedings of the 38th International Conference on Machine Learning (ICML) (PMLR, Vol. 139). 5606–5615. https://arxiv.org/abs/2005.13249

    Pith/arXiv arXiv 2021

  26. [26]

    Gettes, James J

    Paul Kligfield, Leonard S. Gettes, James J. Bailey, Rory Childers, Barbara J. Deal, E. William Hancock, Gerard van Herpen, Jan A. Kors, Peter Macfarlane, David M. Mirvis, Olle Pahlm, Pentti Rautaharju, and Galen S. Wagner. 2007. Recommenda- tions for the Standardization and Interpretation of the Electrocardiogram: Part I.Circulation115, 10 (2007), 1306–13...

    work page doi:10.1161/circulationaha.106 2007

  27. [27]

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pier- son, Been Kim, and Percy Liang. 2020. Concept Bottleneck Models. InProceedings of the 37th International Conference on Machine Learning (ICML). 5338–5348. https://arxiv.org/abs/2007.04612

    Pith/arXiv arXiv 2020

  28. [28]

    Lundberg and Su-In Lee

    Scott M. Lundberg and Su-In Lee. 2017. A Unified Approach to Interpreting Model Predictions. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 30. 4765–4774. https://arxiv.org/abs/1705.07874

    Pith/arXiv arXiv 2017

  29. [29]

    Alireza Makhzani and Brendan Frey. 2013. k-Sparse Autoencoders.arXiv preprint arXiv:1312.5663(2013). https://arxiv.org/abs/1312.5663

    Pith/arXiv arXiv 2013

  30. [30]

    Lau, Jan C

    Dominique Makowski, Tam Pham, Zen J. Lau, Jan C. Brammer, François Lespinasse, Hung Pham, Christopher Schölzel, and S. H. Annabel Chen. 2021. NeuroKit2: A Python Toolbox for Neurophysiological Signal Processing.Behavior Research Methods53, 4 (2021), 1689–1696. doi:10.3758/s13428-020-01516-y

    doi:10.3758/s13428-020-01516-y 2021

  31. [31]

    Kaden McKeen, Laura Oliva, Sameer Masood, et al . 2024. ECG-FM: An Open Electrocardiogram Foundation Model.arXiv preprint arXiv:2408.05178(2024). https://arxiv.org/abs/2408.05178 Published in JAMIA Open, 2025

    Pith/arXiv arXiv 2024

  32. [32]

    Temesgen Mehari and Nils Strodthoff. 2022. Self-Supervised Representation Learning from 12-Lead ECG Data.Computers in Biology and Medicine141 (2022), 105114. doi:10.1016/j.compbiomed.2021.105114

    arXiv 2022

  33. [33]

    Yeongyeon Na, Minje Park, Yunwon Tae, and Sunghoon Joo. 2024. Guiding Masked Representation Learning to Capture Spatio-Temporal Relationship of Electrocardiogram. InInternational Conference on Learning Representations (ICLR). https://arxiv.org/abs/2402.09450

    Pith/arXiv arXiv 2024

  34. [34]

    Olshausen and David J

    Bruno A. Olshausen and David J. Field. 1996. Emergence of Simple-Cell Receptive Field Properties by Learning a Sparse Code for Natural Images.Nature381, 6583 (1996), 607–609. doi:10.1038/381607a0

    doi:10.1038/381607a0 1996

  35. [35]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 32. 8024–8035. https://arxiv.org/abs/1912.01703

    Pith/arXiv arXiv 2019

  36. [36]

    Fabian Pedregosa, Gaël Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and Édouard Duchesnay. 2011. Scikit-learn: Machine Learning in Python.Journal of Machine Learning Re...

    2011

  37. [37]

    Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, János Kramár, Rohin Shah, and Neel Nanda. 2024. Improving Dictionary Learning with Gated Sparse Autoencoders.arXiv preprint arXiv:2404.16014(2024). KDD ’27, August 1–5, 2027, San Jose, CA, USA Duan et al. https://arxiv.org/abs/2404.16014

    Pith/arXiv arXiv 2024

  38. [38]

    Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, János Kramár, and Neel Nanda. 2024. Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders.arXiv preprint arXiv:2407.14435(2024). https://arxiv.org/abs/2407.14435

    Pith/arXiv arXiv 2024

  39. [39]

    Rautaharju, Borys Surawicz, and Leonard S

    Pentti M. Rautaharju, Borys Surawicz, and Leonard S. Gettes. 2009. AHA/AC- CF/HRS Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part IV: The ST Segment, T and U Waves, and the QT In- terval.Journal of the American College of Cardiology53, 11 (2009), 982–991. doi:10.1016/j.jacc.2008.12.014

    work page doi:10.1016/j.jacc.2008.12.014 2009

  40. [40]

    Ribeiro, Manoel Horta Ribeiro, Gabriela M

    Antonio H. Ribeiro, Manoel Horta Ribeiro, Gabriela M. M. Paixão, Derick M. Oliveira, Paulo R. Gomes, Jéssica A. Canazart, Milton P. S. Ferreira, Carl R. Ander- sson, Peter W. Macfarlane, Wagner Meira Jr, Thomas B. Schön, and Antonio Luiz P. Ribeiro. 2020. Automatic Diagnosis of the 12-Lead ECG Using a Deep Neural Net- work.Nature Communications11, 1 (2020...

    doi:10.1038/s41467-020-15432-4 2020

  41. [41]

    Salerno, Patrick C

    Stephen M. Salerno, Patrick C. Alguire, and Herbert S. Waxman. 2003. Com- petency in Interpretation of 12-Lead Electrocardiograms: A Summary and Ap- praisal of Published Evidence.Annals of Internal Medicine138, 9 (2003), 751–760. doi:10.7326/0003-4819-138-9-200305060-00013

    work page doi:10.7326/0003-4819-138-9-200305060-00013 2003

  42. [42]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan- tam, Devi Parikh, and Dhruv Batra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan- tam, Devi Parikh, and Dhruv Batra. 2017. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. InProceedings of the IEEE Inter- national Conference on Computer Vision (ICCV). 618–626. doi:10.1109/ICCV.2017. 74

    doi:10.1109/iccv.2017 2017

  43. [43]

    Elana Simon and James Zou. 2025. InterPLM: Discovering Interpretable Features in Protein Language Models via Sparse Autoencoders.Nature Methods22 (2025), 2107–2117. doi:10.1038/s41592-025-02836-7

    doi:10.1038/s41592-025-02836-7 2025

  44. [44]

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic Attribution for Deep Networks. InProceedings of the 34th International Conference on Machine Learning (ICML). 3319–3328. https://arxiv.org/abs/1703.01365

    Pith/arXiv arXiv 2017

  45. [45]

    Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, et al

  46. [46]

    Alexander Matt Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Monte MacDiarmid. 2023. Activation Addition: Steering Language Models Without Optimization.arXiv preprint arXiv:2308.10248(2023). https://arxiv.org/ abs/2308.10248

    Pith/arXiv arXiv 2023

  47. [47]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. InAdvances in Neural Information Processing Systems (NeurIPS), Vol. 30. 5998–6008. https://arxiv.org/abs/1706.03762

    Pith/arXiv arXiv 2017

  48. [48]

    Lunze, Wojciech Samek, and Tobias Schaeffter

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I. Lunze, Wojciech Samek, and Tobias Schaeffter. 2020. PTB-XL, a Large Publicly Available Electrocardiography Dataset.Scientific Data7, 1 (2020), 154. doi:10. 1038/s41597-020-0495-6

    2020

  49. [49]

    short_name

    Dongdong Zhang, Samuel Yang, Xiaohui Yuan, and Ping Zhang. 2021. Inter- pretable Deep Learning for Automatic Diagnosis of 12-Lead Electrocardiogram. iScience24, 4 (2021), 102373. doi:10.1016/j.isci.2021.102373 CADENCE: Interpretable ECG Foundation Models KDD ’27, August 1–5, 2027, San Jose, CA, USA A Additional Results A.1 Hyperparameter selection We sele...

    arXiv 2021

  50. [2022]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Masked Autoencoders Are Scalable Vision Learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 16000– 16009. doi:10.1109/CVPR52688.2022.01553

    arXiv 2022

  51. [2023]

    https://arxiv.org/abs/2309.08600 Published at ICLR 2024

    Sparse Autoencoders Find Highly Interpretable Features in Language Models.arXiv preprint arXiv:2309.08600(2023). https://arxiv.org/abs/2309.08600 Published at ICLR 2024

    Pith/arXiv arXiv 2023

  52. [2024]

    https://transformer-circuits.pub/ 2024/scaling-monosemanticity/

    Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet.Transformer Circuits Thread(2024). https://transformer-circuits.pub/ 2024/scaling-monosemanticity/

    2024

  53. [2025]

    In Proceedings of the 42nd International Conference on Machine Learning (ICML) (PMLR, Vol

    From Mechanistic Interpretability to Mechanistic Biology: Training, Eval- uating, and Interpreting Sparse Autoencoders on Protein Language Models. In Proceedings of the 42nd International Conference on Machine Learning (ICML) (PMLR, Vol. 267). 460–476. doi:10.1101/2025.02.06.636901

    doi:10.1101/2025.02.06.636901 2025