Pith. sign in

REVIEW 2 major objections 4 minor 42 references

ECG foundation models have distinct, reproducible interpretability profiles that no single accuracy-style score predicts; capacity-matched sparse autoencoders can measure them.

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 07:54 UTC pith:TDELFWDY

load-bearing objection A carefully controlled interpretability benchmark for ECG foundation models with a real result (reconstruction vs accessibility leader split); the one load-bearing unvalidated assumption is the depth-alignment rule, which needs an audit before the orderings are treated as settled. the 2 major comments →

arxiv 2607.27404 v1 pith:TDELFWDY submitted 2026-07-29 cs.LG cs.AI

ECG-InterpBench: Benchmarking the Interpretability of ECG Foundation Models with Matched-Scale Sparse Autoencoders

Yixuan Duan , Wei Qiu This is my paper
classification cs.LG cs.AI
keywords ECG foundation modelssparse autoencodersinterpretability benchmarkcapacity matchingreconstruction fidelityclinical concept accessibilitycross-seed reproducibilityPTB-XL and MIMIC-IV-ECG
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 interpretability of ECG foundation-model representations is a measurable, multi-axis property that differs across models and is not predictable from downstream task accuracy. It introduces ECG-InterpBench, which decomposes frozen representations with sparse autoencoders whose dictionary width and sparsity budget are exactly matched across models, at five standardized encoder depths, and reports three separate axes: sparse reconstruction fidelity, how strongly 49 clinical ECG measurements are exposed by a single sparse coordinate, and cross-seed reproducibility of the learned features. The central empirical claim is that these axes identify different leaders—HuBERT-ECG leads reconstruction while ECG-JEPA leads single-coordinate clinical accessibility—and that this split reproduces on an external MIMIC-IV-ECG cohort. A sympathetic reader would care because a shared representation propagates to every downstream clinical model; knowing which representation is faithfully, locally, and reproducibly interpretable is a selection criterion that accuracy numbers alone do not provide.

Core claim

The paper's central claim is that representation interpretability in ECG foundation models separates into complementary axes that can be reliably compared only when the sparse-autoencoder instrument is matched in capacity across all models and depths. Using a 450-cell atlas of BatchTopK sparse autoencoders with identical widths N∈{768, 3072, 6144, 12288, 24576}, an identical active budget k=96, five relative depths, and three seeds, the paper reports distinct metric-specific orderings. On PTB-XL, reconstruction fidelity ranks HuBERT-ECG first (AUC 0.985, rank-1 probability 1.000) while single-coordinate clinical accessibility ranks ECG-JEPA first (semantic AUC 0.351), with the accessibility

What carries the argument

BatchTopK sparse autoencoders used as matched-scale measurement instruments. For every frozen FM representation (audited width d=768 at every layer), an SAE is trained with the same absolute dictionary widths N and the same per-batch active budget k=96, so all models are compared at identical capacity. The second piece is the relative-depth coordinate ℓ(q)=⌊q(L−1)+1/2⌋, which maps each encoder's layers to the same five relative positions (0, 0.25, 0.5, 0.75, 1). The third piece is leakage-controlled evaluation: for each of 49 ECG concepts, the single SAE coordinate is chosen on a fixed training subset and frozen before held-out test. Together these make cross-model, cross-depth, and cross-sc

Load-bearing premise

The load-bearing premise is that the same five relative depths (0, 0.25, 0.5, 0.75, 1) mark functionally comparable stages of processing across encoders that differ in layer count and input interface; if that mapping fails, every depth-averaged comparison in the benchmark is confounded with architecture.

What would settle it

Re-run the five-depth summaries with layers aligned by absolute position from the final layer (or by a functional similarity measure such as representation CKA to input/output) and show that the metric-specific orderings—particularly the semantic AUC ranking—change materially; or show that a randomly initialized encoder produces the same depth–scale surfaces under the identical matched-SA E protocol.

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

If this is right

  • If the paper is right, downstream model selection should treat interpretability as several axes rather than one: the model that reconstructs best can be mid-pack on clinical accessibility and vice versa.
  • The same capacity-matched SAE protocol can be applied to any frozen encoder with a fixed-width representation to produce comparable interpretability profiles, not just ECG models.
  • Depth–scale surfaces rather than single-layer results are the correct unit for locating where clinical concepts become locally accessible, and the paper's 450-cell atlas provides such surfaces.
  • Reconstruction fidelity and dictionary utilization are decoupled: HuBERT-ECG has the highest reconstruction average but also the highest dead-feature fraction, so both must be reported jointly.
  • Sparsity parameterization (fixed k/d vs fixed k/N) changes accessibility and coverage rankings more than reconstruction rankings, implying benchmarks should sweep the active budget rather than fix one value.

Where Pith is reading between the lines

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

  • The depth-alignment assumption is untested: if the same relative-depth coordinate does not mark functionally comparable stages across encoders with different layer counts, pooling, and tokenization, the depth-averaged AUCs may encode architecture artifacts rather than representation quality. A direct test would align layers by a functional similarity measure and re-run the comparisons.
  • The single-coordinate accessibility metric is a generic instrument that could be carried to other medical modalities (EEG, imaging, EHR), but the clinical concept panel must be modality-matched; the paper's 49-concept PTB-XL panel does not transfer as-is.
  • The finding that accessibility and reconstruction identify different leaders suggests the benchmark is measuring at least two latent capacities—localization versus faithful compression—which may have different implications for different clinical deployment targets.
  • Because the calibration ladder shows single SAE coordinates outperform matched random dictionaries for every model, the sparse structure is non-random; a natural next step would be clinician validation of the semantic content of the best-matching coordinates rather than only correlational alignment.

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

2 major / 4 minor

Summary. The paper introduces ECG-InterpBench, a benchmark for comparing the interpretability of six ECG foundation-model representations using capacity-matched sparse autoencoders. The design fixes a common set of SAE dictionary widths, active-code budgets, training seeds, and relative-depth targets, and evaluates each frozen encoder on three axes: sparse reconstruction fidelity, single-coordinate clinical accessibility/coverage over 49 PTB-XL concepts, and cross-seed feature/subspace reproducibility. Results are summarized as common-scale AUCs over the expansion-factor grid (Eq. 7), averaged over five relative depths and three seeds. The headline findings are that HuBERT-ECG has the highest reconstruction fidelity while ECG-JEPA has the highest single-coordinate clinical accessibility, and that a MIMIC-IV-ECG replication preserves this separation of leaders. The paper reports extensive robustness work: paired patient-cluster bootstraps, BH-corrected contrasts, a dense/random calibration ladder, an input-harmonization audit, a sparsity-sensitivity arm, and the release of code and manifests.

Significance. If the claims hold, ECG-InterpBench is a useful and well-disciplined contribution to the evaluation of ECG foundation models. It moves beyond downstream accuracy to a structured, capacity-controlled assessment of representation interpretability, and it provides a concrete protocol—exact layer-width matching, training-only coordinate selection, paired patient bootstraps, and calibration against dense and random readouts—that other modalities could adapt. The separation between reconstruction fidelity and clinical accessibility as distinct, non-interchangeable properties is a genuinely informative finding, and the external MIMIC replication strengthens the descriptive claim. The authors also deserve credit for preregistering the sensitivity arm, fixing seeds, publishing executable code and manifests, and reporting null or unstable cells rather than cherry-picking favorable capacities.

major comments (2)
  1. [§3.2, Eq. (1); Tables S1/S2; Figs. 3–4] The depth-averaged model profiles and the leader claims depend on the assumption that relative depth q is a functionally commensurate coordinate across encoders with different layer counts. The paper validates the input interface (Appendix F) but does not provide an analogous audit for depth alignment. Eq. (1) maps q∈{0,0.25,0.5,0.75,1} to layers (0,1,3,4,5) for CSFM’s 6 states and to (0,3,6,8,11) for the 12-state encoders; all common-scale AUCs in Table S1, Fig. 3, Fig. 4, and Table 2 then average over these five depths. If, for example, CSFM’s q=0.5 layer is not the processing-stage analogue of ECG-JEPA’s q=0.5 layer, the depth-averaged ordering is confounded by architecture. This is a structural issue for the primary metric, not a secondary one. I request either (i) per-depth model orderings to show that the leader claims are not an artifact of the alignment, (ii) an alignment audit a
  2. [Abstract; §4.3; Appendix E] The abstract and §4.3 describe the MIMIC-IV-ECG study as a 'matched replication' and a 'complete 450-cell multiscale SAE audit,' but Appendix E states that the semantic-alignment and coverage axes use only seven continuous waveform concepts, whereas the PTB-XL benchmark uses 49 concepts (Section 3.1, §4.2). The 450 SAE cells are matched, but the clinical-accessibility measurement is not concept-matched. Since the central claim is that the replication confirms the separation of reconstruction and accessibility leaders, a 7-concept panel is a substantially weaker test, especially for coverage. Please either qualify the 'matched replication' wording in the abstract, justify that seven concepts are sufficient (e.g., by showing leader stability across concept families on the PTB-XL 49-concept panel), or repeat the full 49-concept evaluation on the MIMIC cohort.
minor comments (4)
  1. [Table S1] The model column contains a typo: 'CSFMe' should be 'CSFM'.
  2. [§4.4, Table S2] The text states that the full SAE retains 79.9–87.4% of dense-ridge mean |r|, but computing Full SAE / Dense ridge from Table S2 gives a range of roughly 82–89% (e.g., 0.677/0.764 ≈ 0.886 for CARDIAC-FM and 0.536/0.651 ≈ 0.823 for CSFM). Please check the arithmetic or clarify the denominator used.
  3. [§3.4 and Appendix B] The symbol α in 'fixed-α=10 ridge' is undefined; if it is the L2 regularization weight, please state so explicitly.
  4. [Figure 2 and Figure 4 labels] The y-axis label 'semantic AUC' is used interchangeably with 'single-coordinate clinical accessibility AUC' in the text. Since Eq. (7) defines a generic AUC, please use consistent terminology (e.g., 'accessibility AUC') to avoid confusion with any semantic-similarity measure.

Circularity Check

0 steps flagged

No significant circularity: all reported metrics are held-out measurements under explicit operational definitions, with no fitted parameter presented as a prediction and no load-bearing self-citation chain.

full rationale

ECG-InterpBench is an empirical benchmarking paper rather than a derivation of predictions from first principles. Its central quantities are defined operationally and then measured on held-out data. The accessibility score uses Eq. (6) to select a coordinate on a fixed training subset and then evaluates the same coordinate on validation/test records; this is standard held-out feature selection, not a fitted parameter being relabeled as a prediction. Eq. (7) is an explicit aggregation rule (normalized log-scale area under a measured curve), not a result that re-imports its own conclusion. Eq. (1) maps relative depth to layer indices; this is a design choice for cross-encoder comparison. If relative depth is not functionally commensurate across architectures, the depth-averaged comparisons would be confounded, but that is a validity/assumption concern, not circularity by construction: the paper does not define 'depth' in terms of the ordering it later reports, nor does any equation force the reported model rankings. There are no self-citations by the authors, and no uniqueness theorem from prior work is imported to justify a choice. The MIMIC-IV-ECG replication uses the same protocol on a different cohort and is therefore an external check, not a circular confirmation. The paper's own Limitations section explicitly disclaims mechanism, clinical utility, and prospective outcomes, which further shows the claims are bounded empirical measurements rather than tautological derivations. The input-harmonization audit in Appendix F addresses input confounding but does not create a circular step. Overall, the derivation chain is self-contained as a measurement protocol: every reported AUC is an integral of observed, held-out quantities, and no load-bearing conclusion reduces by definition to its inputs.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The measurements rest on hand-chosen grids (E, q, k/d), hand-chosen thresholds (|r| ≥ 0.20, R² ≥ 0.90, dead < 0.20), and a borrowed concept panel; the SAE-as-instrument validity is imported from the LM interpretability literature rather than re-established for ECG. No new entities are postulated — the 'interpretability atlas' and 'semantic AUC' are measurement constructions, not additions to the ontology. The relative-depth coordinate (Eq. 1) is the most paper-specific axiom and is unaudited. Sensitivity arms (§3.7, Appendix D) bound some but not all of these dependencies.

free parameters (6)
  • Active budget k/d = 1/8 (k = 96) = k=96 at all widths (d=768)
    Hand-chosen sparsity protocol for the primary arm; the fixed-k/N sensitivity arm (§3.7) checks robustness but the primary ordering is defined under this budget.
  • Expansion grid E = N/d ∈ {1, 4, 8, 16, 32}
    Hand-chosen widths; the common-scale AUC (Eq. 7) integrates over exactly this support, so all model-level summaries are conditional on the grid.
  • Relative-depth targets q ∈ {0, .25, .5, .75, 1}
    Hand-chosen depth positions mapped by Eq. (1); every model-level AUC averages over these five targets, making results conditional on the depth alignment.
  • Coverage threshold |r| ≥ 0.20 = 0.20
    Hand-chosen 'consistent descriptive operating point' (§3.3); coverage ordering responds to this cutoff (Kendall τ min 0.47–0.60 across E).
  • Validation fidelity gate (R² ≥ 0.90, dead fraction < 0.20)
    Hand-chosen thresholds used to categorize layer–scale cells (§3.5); not swept, so which cells count as 'passing' depends on them.
  • SAE training schedule (8,000 steps, batch 256, Adam LR 3e-4, seeds 4311–4313)
    Standard settings chosen by hand; the exact values affect reported reconstruction R² and stability levels across the atlas.
axioms (5)
  • domain assumption BatchTopK SAEs with unit-norm decoder columns produce valid sparse decompositions of frozen transformer representations (refs 6, 8, 9, 22, 35).
    The benchmark's validity as an interpretability instrument is imported from the LM–SAE literature; the paper does not independently establish that SAE directions on ECG states are semantically meaningful beyond beating random dictionaries.
  • ad hoc to paper Relative depth ℓ(q)=⌊q(L−1)+1/2⌋ is a commensurate coordinate across encoders with different layer counts (Eq. 1, §3.2).
    Design choice specific to this paper; no validation that equal relative positions implement functionally comparable stages across 6- and 12-layer encoders.
  • domain assumption The 49 PTB-XL waveform-derived measurements form a valid, consistent clinical concept panel (§3.1, §3.3).
    Concept list is borrowed from PTB-XL; clinical relevance is assumed, and the MIMIC audit swaps in only 7 of these concepts, breaking panel consistency across cohorts.
  • domain assumption Validation reconstruction R² and dead-feature fraction measure sparse fidelity (§3.5).
    Standard SAE diagnostics; treated as adequate proxies for decomposition faithfulness of the sparse instrument.
  • standard math Train-only argmax selection over up to 24,576 coordinates, frozen before test evaluation, controls leakage (§3.3).
    Standard train/select/evaluate protocol; held-out correlation levels (mean |r| ≈ 0.26–0.35) empirically bound the selection bias.

pith-pipeline@v1.3.0-daily-deepseek · 17687 in / 23462 out tokens · 225628 ms · 2026-08-01T07:54:54.087287+00:00 · methodology

0 comments
read the original abstract

Existing benchmarks for electrocardiogram foundation models primarily evaluate downstream predictive performance, providing limited insight into whether their internal representations can be faithfully decomposed, clinically interpreted, or reproduced across independent analyses. We introduce ECG-InterpBench, a benchmark designed to systematically evaluate the interpretability of ECG foundation-model representations. ECG-InterpBench uses sparse autoencoders as standardized measurement instruments and matches their capacity across models to enable controlled comparisons. We evaluate six frozen ECG foundation models across five standardized encoder depths, five matched dictionary widths, and three random seeds, producing a 450-cell interpretability atlas comprising 75 exactly matched six-model comparison blocks. The benchmark evaluates complementary dimensions of representation interpretability, including sparse reconstruction fidelity, single-feature accessibility and coverage of 49 clinically meaningful ECG measurements, and cross-seed feature reproducibility. The evaluation further quantifies patient-sampling uncertainty, depth- and seed-dependent variation, and sensitivity to the sparsity parameterization. The benchmark reveals that ECG foundation models exhibit distinct interpretability profiles. A matched replication on MIMIC-IV-ECG confirms that reconstruction fidelity and clinical accessibility identify different leading models. The benchmark is accompanied by executable evaluation code, standardized manifests, cell-level metrics, and reproducibility audits. ECG-InterpBench complements performance-centered ECG benchmarks by providing a capacity-controlled and reproducible framework for comparing ECG foundation models across distinct dimensions of representation interpretability.

Figures

Figures reproduced from arXiv: 2607.27404 by Wei Qiu, Yixuan Duan.

Figure 1
Figure 1. Figure 1: Overview of ECG-InterpBench. (a) PTB-XL and credentialed MIMIC-IV ECGs are passed through six frozen ECG [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Single-coordinate clinical accessibility across the matched SAE atlas. (a) Test mean absolute concept correlation over [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Patient-bootstrap FM profiles at each common SAE scale: (a) reconstruction [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Metric-specific common-scale AUC estimates and paired patient-cluster intervals for (a) reconstruction, (b) single [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Robustness axes. (a) Cross-seed decoder matching above random and (b) selected-subspace overlap over common scale. [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

42 extracted references · 1 canonical work pages

  1. [1]

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. 2018. Sanity Checks for Saliency Maps. InAdvances in Neural Information Processing Systems, Vol. 31. https://arxiv.org/abs/1810.03292

    Pith/arXiv arXiv 2018

  2. [2]

    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

  3. [3]

    Nora Belrose, David Schneider-Joseph, Shauli Ravfogel, Ryan Cotterell, Edward Raff, and Stella Biderman. 2023. LEACE: Perfect Linear Concept Erasure in Closed Form. InAdvances in Neural Information Processing Systems, Vol. 36. https://arxiv.org/abs/2306.03819

    Pith/arXiv arXiv 2023

  4. [4]

    Yoav Benjamini and Yosef Hochberg. 1995. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing.Journal of the Royal Statistical Society: Series B (Methodological)57, 1 (1995), 289–300. doi:10.1111/j. 2517-6161.1995.tb02031.x

    arXiv 1995

  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]

    Edoardo Coppola, Mattia Savardi, Mauro Massussi, Marianna Adamo, Marco Metra, and Alberto Signoroni. 2024. HuBERT-ECG as a Self-Supervised Foun- dation Model for Broad and Scalable Cardiac Applications.medRxiv(2024). doi:10.1101/2024.11.14.24317328

    doi:10.1101/2024.11.14.24317328 2024

  8. [8]

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

  9. [9]

    Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2025. Scaling and Evaluating Sparse Autoencoders. InInternational Conference on Learning Representations. https://openreview.net/forum?id=tcsZt9ZNKD

    2025

  10. [10]

    Marzyeh Ghassemi, Luke Oakden-Rayner, and Andrew L. Beam. 2021. The False Hope of Current Approaches to Explainable Artificial Intelligence in Health Care.The Lancet Digital Health3, 11 (2021), e745–e750. doi:10.1016/S2589- 7500(21)00208-9

    doi:10.1016/s2589- 2021

  11. [11]

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

    Brian Gow, Tom Pollard, Larry A. Nathanson, Alistair Johnson, Benjamin Moody, Catarina Fernandes, Nathaniel Greenbaum, Jonathan W. Waks, Parastou Eslami, Timothy Carbonati, et al. 2023.MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset. doi:10.13026/4nqg-sb35

    doi:10.13026/4nqg-sb35 2023

  12. [12]

    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, et al. 2026. Cardiac Health Assessment across Scenarios and Devices Using a Multimodal Founda- tion Model Pretrained on Data from 1.7 Million Individuals.Nature Machine Intelligence8, 2 (2026), 220–233. doi:10.1038/s422...

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

  13. [13]

    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 (2019), 65–69. doi:10.1038/ s41591-018-0268-3

    2019

  14. [14]

    Neil Jethani, Mukund Sudarshan, Yindalon Aphinyanaphongs, and Rajesh Ran- ganath. 2023. Have We Learned to Explain?: How Interpretability Methods Can Learn to Encode Predictions in Their Interpretations. InProceedings of the 26th International Conference on Artificial Intelligence and Statistics (Proceedings of Machine Learning Research, Vol. 206). 1459–1...

    2023

  15. [15]

    Adam Karvonen, Can Rager, Johnny Lin, Curt Tigges, Joseph Bloom, David Chanin, Yeu-Tong Lau, Eoin Farrell, Callum McDougall, Kola Ayonrinde, Matthew Wearden, Arthur Conmy, Samuel Marks, and Neel Nanda. 2025. SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability.arXiv preprint arXiv:2503.09532(2025). https://arxiv.o...

    Pith/arXiv arXiv 2025

  16. [16]

    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 (Proceedings of Machine Learning Research, Vol. 80). 2668–2677. https:...

    2018

  17. [17]

    Sehun Kim. 2024. Learning General Representation of 12-Lead Electrocardiogram with a Joint-Embedding Predictive Architecture.arXiv preprint arXiv:2410.08559 (2024). https://arxiv.org/abs/2410.08559

    Pith/arXiv arXiv 2024

  18. [18]

    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 (Proceedings of Machine Learning Research, Vol. 139). 5606–5615. https://proceedings.mlr.press/v139/ kiyasseh21a.html

    2021

  19. [19]

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. 2020. Concept Bottleneck Models. In Proceedings of the 37th International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 119). 5338–5348. https://proceedings.mlr. press/v119/koh20a.html

    2020

  20. [20]

    William Lehn-Schiøler, Magnus Ruud Kjær, Rahul Thapa, Magnus Guldberg Pedersen, Anton Storgaard Mosquera, Nick Williams, Radu Gatej, Tue Lehn- Schiøler, Andreas Brink-Kjær, Sadasivan Puthusserypady, Sándor Beniczky, James Zou, and Lars Kai Hansen. 2026. Mechanistic Interpretability of EEG Foundation Models via Sparse Autoencoders.arXiv preprint arXiv:2605...

    Pith/arXiv arXiv 2026

  21. [21]

    Brody, Vidhushei Yogeswaran, Kerri L

    Fumin Li, Siting Li, Yuhan Qian, Bojun Chen, Jennifer A. Brody, Vidhushei Yogeswaran, Kerri L. Wiggins, Colleen M. Sitlani, Joshua C. Bis, Ali Shojaie, et al

  22. [22]

    Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, János Kramár, Anca Dragan, Rohin Shah, and Neel Nanda. 2024. Gemma Scope: Open Sparse Autoencoders Everywhere All at Once on Gemma 2.arXiv preprint arXiv:2408.05147(2024). https://arxiv.org/abs/2408. 05147

    Pith/arXiv arXiv 2024

  23. [23]

    Julien Mairal, Francis Bach, Jean Ponce, and Guillermo Sapiro. 2009. Online Dictionary Learning for Sparse Coding. InProceedings of the 26th International Conference on Machine Learning. 689–696. doi:10.1145/1553374.1553463

    arXiv 2009

  24. [24]

    Aleksandar Makelov, Georg Lange, and Neel Nanda. 2025. Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control. InInterna- tional Conference on Learning Representations. https://openreview.net/forum? id=1Njl73JKjB

    2025

  25. [25]

    Kaden McKeen et al. 2025. ECG-FM: An Open Electrocardiogram Foundation Model.JAMIA Open8, 5 (2025), ooaf122. doi:10.1093/jamiaopen/ooaf122

    doi:10.1093/jamiaopen/ooaf122 2025

  26. [26]

    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

  27. [27]

    James Murdoch, Chandan Singh, Karl Kumbier, Reza Abbasi-Asl, and Bin Yu

    W. James Murdoch, Chandan Singh, Karl Kumbier, Reza Abbasi-Asl, and Bin Yu

  28. [28]

    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. https: //openreview.net/forum?id=WcOohbsF4H

    2024

  29. [29]

    Olshausen and David J

    Bruno A. Olshausen and David J. Field. 1997. Sparse Coding with an Over- complete Basis Set: A Strategy Employed by V1?Vision Research37, 23 (1997), 3311–3325. doi:10.1016/S0042-6989(97)00169-7

    doi:10.1016/s0042-6989(97)00169-7 1997

  30. [30]

    Ribeiro, Manoel Horta Ribeiro, Gabriela M

    Antônio H. Ribeiro, Manoel Horta Ribeiro, Gabriela M. M. Paixão, Derick M. Oliveira, Paulo R. Gomes, Jéssica A. Canazart, Milton P. S. Ferreira, Carl R. An- dersson, 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 Network.Nature Communications11 (2020), ...

    doi:10.1038/s41467- 2020

  31. [31]

    Cynthia Rudin. 2019. Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead.Nature Machine Intelligence1 (2019), 206–215. doi:10.1038/s42256-019-0048-x

    doi:10.1038/s42256-019-0048-x 2019

  32. [32]

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

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

  33. [33]

    Nils Strodthoff and Claas Strodthoff. 2019. Detecting and Interpreting Myocardial Infarction Using Fully Convolutional Neural Networks.Physiological Measure- ment40, 1 (2019), 015001. doi:10.1088/1361-6579/aaf34d

    doi:10.1088/1361-6579/aaf34d 2019

  34. [34]

    Nils Strodthoff, Patrick Wagner, Tobias Schaeffter, and Wojciech Samek. 2021. Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL.IEEE Journal of Biomedical and Health Informatics25, 5 (2021), 1519–1528. doi:10.1109/ JBHI.2020.3022989

    arXiv 2021

  35. [35]

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

    2024

  36. [36]

    McCradden, and Anna Goldenberg

    Sana Tonekaboni, Shalmali Joshi, Melissa D. McCradden, and Anna Goldenberg

  37. [37]

    Akhil Vaid et al. 2023. A Foundational Vision Transformer Improves Diagnostic Performance for Electrocardiograms.npj Digital Medicine6 (2023), 108. doi:10. Duan and Qiu 1038/s41746-023-00840-9

    2023

  38. [38]

    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 (2020), 154. doi:10.1038/ s41597-020-0495-6 ECG-InterpBench Table S1: Common-five-scale model profiles with paired patient-cluster 95% inter...

    arXiv 2020

  39. [40]

    InProceedings of the 4th Machine Learning for Healthcare Conference (Proceedings of Machine Learning Research, Vol

    What Clinicians Want: Contextualizing Explainable Machine Learning for Clinical End Use. InProceedings of the 4th Machine Learning for Healthcare Conference (Proceedings of Machine Learning Research, Vol. 106). 359–380. https: //proceedings.mlr.press/v106/tonekaboni19a.html

  40. [2019]

    Proceedings of the National Academy of Sciences116, 44 (2019), 22071–22080

    Definitions, Methods, and Applications in Interpretable Machine Learning. Proceedings of the National Academy of Sciences116, 44 (2019), 22071–22080. doi:10.1073/pnas.1900654116

    doi:10.1073/pnas.1900654116 2019

  41. [2024]

    InInternational Conference on Learning Representations

    Sparse Autoencoders Find Highly Interpretable Features in Language Mod- els. InInternational Conference on Learning Representations. https://openreview. net/forum?id=F76bwRSLeK

  42. [2026]

    doi:10.64898/2026.03

    CARDIAC-FM: A Multimodal Foundation Model for Cardiovascular Risk Prediction Using ECG and Cardiac MRI.medRxiv(2026). doi:10.64898/2026.03. 16.26348526

    work page doi:10.64898/2026.03 2026