Pith. sign in

REVIEW 3 major objections 4 minor 51 references

CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read CardioState-JEPA claims that ECG, PPG, and PCG are time-delayed renderings of one shared cardiac state, and that a single frozen encoder trained with masked latent prediction plus delay-aware cross-modal prediction can learn that state…

desk verdict A thoughtful shared-cardiac-encoder paper whose PPG results may be contaminated by pretraining/evaluation overlap; worth peer review but needs the overlap ruled out. read the letter →

arxiv 2608.12944 v1 pith:N4YZ3OIL submitted 2026-08-13 cs.LG eess.IVstat.ML

Hamza Shafiq , Hung Manh Pham , Bin Zhu , Pan Zhou , Jun Hu , Aaqib Saeed This is my paper · ORCID
classification cs.LGeess.IVstat.ML
keywords cardiacfoundationmodeljoint-embeddingpredictivearchitectureself-supervisedlearningECGPPGPCGcross-modalalignmentdelay-awarepretraining
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that ECG, PPG, and PCG are not three separate signal families but three time-delayed renderings of one hidden cardiac state, and that a single model can learn that state. It introduces CardioState-JEPA, a two-stage joint-embedding predictive architecture that first performs masked latent prediction on abundant unimodal data and then delay-aware cross-modal prediction on scarce paired data. Evaluated as a frozen encoder on 25 downstream tasks, the shared encoder raises average PPG classification by 8.2 AUROC points, PCG murmur detection by 18.8 points, and ECG classification by 15.5 points over the best self-supervised single-signal baseline. If correct, this makes heterogeneous cardiac signals mutually supervising sources for one cardiac foundation model, reducing the need for labeled or clinically annotated ECG data.

What carries the argument

The load-bearing object is the latent cardiac state $c(t)$ of Eq. (1), with each modality modeled as a rendering $x_m(t)=O_m(c(t-\tau_m(t)), u_m(t))$. The carrying mechanism is delay-aware cross-modal prediction: a learned delay head estimates per-token offsets $\tau_{m\to n}$, a Gaussian kernel gathers the target token at the shifted time, and the shared predictor matches source predictions to those aligned targets, with delay supervision anchored to R-peak-to-S1 and pulse-arrival-time intervals. The shared predictor and momentum encoder keep both prediction objectives in one space. This machinery forces the encoder to remove sensor-specific appearance and temporal offset so that downstream linear probes read shared cardiac physiology rather than waveform morphology.

What would settle it

Retrain Stage II on paired recordings whose cross-modal pairs have been shuffled so the target signal comes from a different patient while delay supervision still uses the source-only anchor; if the reported downstream gains persist, the cross-modal objective is not learning shared physiology, and if they collapse, the shared-state reading is supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a single shared Transformer pretrained with intra-modal masked latent prediction followed by delay-aware cross-modal prediction yields a cardiac code that is largely invariant to sensing modality while remaining discriminative for downstream tasks. After Stage I, codes cluster by sensor; after Stage II the modality silhouette falls from 0.121 to -0.006, and downstream performance improves. The same frozen encoder attains an average PPG classification AUROC of 80.4, an average ECG AUROC of 84.1 across 18 settings, and 97.9 and 66.8 AUROC on CirCor murmur and CinC2016 abnormal heart sound detection. The authors read this as evidence that the model has aligned ECG, PPG, and PCG in latent cardiac time rather than by raw timestamp.

Load-bearing premise

The load-bearing premise is that ECG, PPG, and PCG are time-delayed renderings of one shared latent cardiac state that a scalar per-token delay can align; if no such common state is recoverable after tokenization, the cross-modal gains would reflect matched sensor appearance rather than shared cardiac physiology.

Editorial extensions

If this is right

  • A single frozen encoder can serve ECG, PPG, and PCG tasks, so one cardiac foundation model can replace three per-sensor models.
  • Signal-only cardiac pretraining can approach or match models trained with clinical text or supervised labels on several ECG benchmarks, lowering the need for privileged annotations.
  • The learned delay aligner tracks physiological timing, so per-token delay estimates are interpretable cardiac timing measurements rather than opaque parameters.
  • The two-stage curriculum matches the practical data regime: scarce synchronized multi-sensor recordings are reserved for alignment while abundant unimodal corpora carry most of the representation learning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper's claims: if the shared-latent view is correct, the delay estimates themselves become a free physiological measurement, so pulse transit time and electromechanical delay could be read directly from a single ECG-PPG or ECG-PCG pair.
  • The same two-stage recipe could transfer to other synchronized physiological families, such as impedance cardiography, ballistocardiography, or fetal monitoring, whenever a beat-level anchor can be detected.
  • A practical extension would align a new sensor's tokens to the existing shared code, letting a rare modality inherit the representation of abundant ones without a large paired corpus.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. CardioState-JEPA proposes a two-stage joint-embedding predictive architecture that learns a single shared cardiac representation across ECG, PPG, and PCG. Stage I performs intra-modal masked latent prediction on large unimodal corpora (MIMIC-IV-ECG, PPG-EXT, BMD-HS); Stage II uses paired recordings for delay-aware cross-modal prediction with a learned, physiology-anchored delay aligner. The frozen shared encoder is then evaluated via linear probing on 25 downstream tasks, reporting average AUROC gains of 8.2 points for PPG classification, 18.8 points for PCG murmur detection, and 15.5 points for ECG classification over the best self-supervised signal baselines. The paper includes ablations over pretraining modalities, learning objectives, and auxiliary losses, as well as sensitivity analyses for loss weights, and visualizes the learned delay alignment and the shared latent space.

Significance. If the reported results hold, the paper makes a substantial contribution by demonstrating that heterogeneous cardiac signals can mutually supervise a single foundation model, a direction that differs from the current per-modality cardiac foundation models. Strengths include the two-stage curriculum, the explicit supervision of the delay aligner with independent physiological anchors (R-peak-to-S1 and pulse arrival time), the breadth of downstream evaluation (17 PPG tasks, 18 ECG settings, 2 PCG tasks), and the detailed ablations and loss-weight sensitivity analyses. The principal caveat is the potential pretraining/evaluation overlap for PPG, which directly affects a headline claim; the ECG and PCG results rest on separate corpora and are less exposed. The absence of released code/checkpoints and of seed-level variance for most tables are secondary but important reproducibility concerns.

major comments (3)
  1. [Section 4.1, Table 1, and A.6] The paper does not establish disjointness between the PPG pretraining corpus (PPG-EXT, 4.6M recordings) and the PPG downstream tasks obtained from PulseLM (e.g., MIMIC AF, PPG Arrhythmia, BIDMC RR, UQVital). The statement in A.6 that downstream segments are 'never seen during pretraining' is an assertion without supporting evidence, and the 'patient-disjoint splits' in Section 4.2 refer only to the linear-probe train/test split. If the same recordings or patients appear in both, the reported 8.2-point average PPG classification gain and the regression MAE improvement (10.9 to 9.1) in Table 1 reflect memorization rather than cross-modal transfer. Please either provide explicit evidence of non-overlap (e.g., recording IDs, patient IDs, or a formal exclusion protocol) or repeat the PPG evaluation with a pretraining set that provably excludes the downstream corpora.
  2. [Tables 1, 2, and 4] Only the PCG results in Table 3 report mean±std over three seeds; the headline PPG and ECG numbers are single values. Since the protocol is a frozen encoder plus a linear probe, rerunning with multiple seeds is computationally inexpensive, and without variance estimates the reader cannot judge whether differences such as PPG Arrhythmia 96.8 vs. 95.8, or the ECG gap between CardioState-JEPA and MoCo-v3, are significant. Please report mean±std (or confidence intervals) over at least three seeds for all reported numbers, or clearly state that the table entries are single runs and include a statistical significance analysis for the main comparisons.
  3. [Section 4.5 and A.9] The evidence that cross-modal training produces a modality-invariant yet task-relevant shared space rests on t-SNE plots and silhouette values computed on 'held-out co-recorded' samples, but the paper does not specify which paired dataset these samples come from, how the held-out split is constructed, or the number of samples, and it reports no variance for the silhouette values. Because SensSmartTech, VitalDB, and EPHNOGRAM are used in Stage II pretraining, the held-out samples must be provably disjoint from the pretraining paired data for the silhouette claim to be meaningful. Please specify the dataset, split, and sample size, and consider adding a quantitative cross-modal correspondence test (e.g., cross-modal retrieval or k-NN classification of one modality against another) to directly support the shared-representation claim.
minor comments (4)
  1. [A.5 and Eqs. (3)-(4)] The sensitivity analysis varies only lambda_cross, lambda_delay, and lambda_state; tau_max and sigma are free parameters that are not covered. Please include a sensitivity analysis for these or justify the chosen values.
  2. [Section 3.4 and Eq. (2)] The symbol H is overloaded: it denotes the encoded token sequences (e.g., H_m) and also the Smooth-L1 loss in Eq. (2). Please use a different symbol for the loss (e.g., L_huber) to avoid ambiguity.
  3. [Section 3.5 and Eq. (4)] The alignment kernel in Eq. (4) uses a softmax over 'in-band target tokens,' but the paper does not define how the in-band set is determined. Please specify the window size or selection rule.
  4. [Figure 3 and Section 4.5] The caption reports silhouette values of 'sil + 0.12' and 'sil -0.01' while the text reports 0.121 and -0.006; please make the numbers consistent.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor non-load-bearing circularity in the delay-alignment check; the central cross-modal transfer claims are externally benchmarked and remain self-contained.

  1. fitted input called prediction [Appendix A.8 (Delay Alignment Visualization) vs. Eq. (7) in Section 3.5]
    "To check that the delay aligner captures physiological timing rather than an arbitrary offset, we visualize its output on paired recordings. ... As shown in Figure 8, the predicted time lines up with the systolic upstroke in the PPG and with the first heart sound S1 in the PCG, and it does so consistently from one beat to the next. This close agreement indicates that the learned delay reflects the true electromechanical and pulse transit timing between modalities rather than a value that merely minimizes the training loss."

    Eq. (7) directly regresses the delay estimate onto tau_i^anchor, where the anchor is defined as the R-peak-to-first-heart-sound interval for ECG-PCG and the pulse arrival time for ECG-PPG. The 'true cardiac event' used in A.8 is the same anchor, so on the training distribution the agreement between t_R + tau_hat and the S1/upstroke is enforced by the loss itself; the visualization restates the supervision rather than providing independent evidence that the delay is physiological. The paper does not state that the shown paired recordings are held out from L_delay-sup, so the condition that would make the check non-circular is absent.

full rationale

The central derivation chain is self-contained: pretraining uses intra-modal masked latent prediction (Eq. 2) and delay-aware cross-modal prediction (Eq. 6) on unlabeled waveforms, with the delay supervised by beat-detection anchors that are independent of downstream labels; all headline gains are measured by frozen-encoder linear probes on external benchmarks (PTB-XL, CPSC, CSN, CirCor, CinC, and PulseLM PPG tasks). Self-citations to D-BETA, PulseLM, and StethoLM appear only as background or reference baselines and are not used to construct the result. The only equation-level circular element is the Appendix A.8 delay-alignment check, which confirms the same anchor supervision used to fit the delay; this is not load-bearing for the main transfer claims. Separately, the paper never establishes that PPG-EXT pretraining windows and PulseLM PPG downstream samples are disjoint, so the 8.2-point PPG classification gain may partly reflect pretraining/evaluation overlap rather than cross-modal transfer; that is a data-leakage risk, not an instance of the result being definitionally equal to its input.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central claim rests on the assumed existence of a shared latent cardiac state and on hand-chosen training weights. No physical entity is introduced beyond the abstract latent state, which is not directly falsifiable outside the paper.

free parameters (6)
  • tau_max (delay head bound)
    Bounds the per-token delay in Eq. (3); set by hand and its exact numeric value is not given in the visible text. It controls how far alignment can shift.
  • sigma (alignment kernel width)
    Width of the Gaussian gather kernel in Eq. (4); set by hand and not reported. It must be comparable to token spacing for differentiable soft alignment.
  • lambda_cross = 1.0
    Weight of the cross-modal JEPA loss in Eq. (8); chosen by hand. Appendix A.5 shows PPG classification varies from 77.3 to 80.4 across nearby values.
  • lambda_delay = 1.0
    Weight of the delay supervision loss in Eq. (8); chosen by hand. Sensitivity analysis shows PPG regression moves by roughly one MAE point across nearby values.
  • lambda_state = 0.05
    Weight of the VICReg-style state term in Eq. (8); chosen by hand to balance modalities. Raising it to 0.25 or 1.0 lowers PPG classification.
  • EMA momentum schedule = 0.998 to 0.9999
    Momentum encoder EMA schedule for stable JEPA targets; standard design choice but part of the pretraining protocol.
assumptions (5)
  • domain assumption Shared latent cardiac state c(t) exists and Eq. (1) holds: each sensor is a rendering of the same state shifted by delay tau_m(t).
    This is the core physiological model in Section 3.2; if false, cross-modal prediction has no shared target to learn.
  • domain assumption A per-token scalar delay is sufficient to align modalities (Eqs. 3 and 4).
    The delay head predicts one offset per source token; real electromechanical and pulse-transit delays vary, but the model assumes a scalar correction plus a Gaussian soft gather captures them.
  • domain assumption Large contiguous masking prevents local interpolation and forces recovery of cardiac state (Stage I, Eq. 2).
    Design assumption that masked latent prediction from neighboring cycles yields rhythm and phase structure rather than waveform appearance.
  • domain assumption Paired corpora are synchronized and beat-detection anchors are reliable enough to supervise delay (Eq. 7).
    Relies on VitalDB, EPHNOGRAM, and SensSmartTech synchronization and on R-peak, S1, and pulse-arrival anchors; the paper masks supervision when clean beats are absent, so noisy anchors are a weaker but still load-bearing premise.
  • domain assumption Linear probing of a frozen encoder is a valid measure of representation transfer.
    Standard evaluation choice but narrow; results could differ under fine tuning or larger encoders, which the limitations section acknowledges.
invented entities (1)
  • Shared latent cardiac state c(t)
    purpose: Hidden variable that all three modalities are assumed to render in Eq. (1); the target that the shared encoder is trained to predict.
    Introduced by the formulation rather than measured directly; downstream task accuracy and alignment visualizations are indirect, in-paper evidence. It is a modeling construct, not an observed physiological quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation." pith.science (2026). https://pith.science/paper/N4YZ3OIL

@misc{pith2026260812944,
  author       = {Pith},
  title        = {Pith review of: CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N4YZ3OIL}},
  note         = {Machine review of arXiv:2608.12944}
}
read the original abstract

Electrocardiography (ECG), photoplethysmography (PPG), and phonocardiography (PCG) provide complementary views of the same cardiac cycle, yet existing cardiac foundation models are trained for a single sensing modality, leaving the shared physiology across sensors unexploited. We introduce CardioState-JEPA, a cardiac foundation model to learn a single shared representation jointly across ECG, PPG, and PCG, built on a physiology-aware joint-embedding predictive architecture. The model maps heterogeneous waveforms into a common token space, processes them with a single shared Transformer encoder, and learns by predicting masked latent cardiac states, placing the pretraining target on shared physiology rather than sensor-specific waveform appearance. To handle the temporal offsets between electrical, mechanical, and hemodynamic events, cross-modal prediction uses a learned delay aligner that matches signals at the corresponding cardiac time. Because synchronized multi-sensor recordings are scarce, CardioState-JEPA first learns within-modality structure from abundant unimodal data and then uses paired data to align modalities in latent cardiac time. Evaluated as a frozen encoder across 25 downstream tasks spanning ECG, PPG, and PCG, our encoder improves average PPG classification by 8.2 AUROC points, PCG murmur detection by 18.8 AUROC points, and ECG classification by 15.5 AUROC points over the best self-supervised signal baseline and matches or exceeds cardiac models trained with privileged clinical text or supervised labels on several ECG benchmarks. These results establish that heterogeneous cardiac signals can mutually supervise a single foundation model of cardiac physiology.

Figures

Figures reproduced from arXiv: 2608.12944 by the authors.

Figure 1
Figure 1. Motivation of CardioState-JEPA. ECG, PPG, and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of CardioState-JEPA. Stage I learns unimodal cardiac codes with masked latent prediction from abundant [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. T-SNE of cardiac codes from co-recorded ECG, PPG, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Synchronized paired and trimodal recordings used [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 7
Figure 7. Figure 7: Attention visualizations from CardioState-JEPA on [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 6
Figure 6. Figure 6: Downstream evaluation signals, shown one modality [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 9
Figure 9. Figure 9: t-SNE of PPG features on the binary atrial fibrillation [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: t-SNE of PPG features on the six-class arrhythmia [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 26 canonical work pages

  1. [1]

    ShamsNafisaAli,AfiaZahin,SamiulBasedShuvo,NusratBinta Nizam, Shoyad Ibn Sabur Khan Nuhash, Sayeed Sajjad Razin, S. M. Sakeef Sani, Farihin Rahman, Nawshad Binta Nizam, Farhat Binte Azam, Rakib Hossen, Sumaiya Ohab, Nawsabah Noor, and Taufiq Hasan. BUET multi-disease heart sound dataset: A comprehensive auscultation dataset for developing computer-aided di...

    work page 2026

  2. [2]

    Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama S. Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wil- son, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: Learning the language of...

    work page 2024

  3. [3]

    Chronos- 2: From univariate to universal forecasting.arXiv preprint arXiv:2510.15821, 2025

    Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, et al. Chronos- 2: From univariate to universal forecasting.arXiv preprint arXiv:2510.15821, 2025

    arXiv

  4. [4]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bo- janowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 15619–15629, 2023. doi: 10.1109/CVPR52729.2023.01499

    arXiv 2023

  5. [5]

    InInternational Conference on Learning Representa- tions (ICLR), 2022

    AdrienBardes,JeanPonce,andYannLeCun.VICReg:Variance- invariance-covariance regularization for self-supervised learn- ing. InInternational Conference on Learning Representa- tions (ICLR), 2022. URL https://openreview.net/forum?id= xm6YD62D1Ub

    work page 2022

  6. [6]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geof- frey Hinton. A simple framework for contrastive learning of visual representations. InProceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings ofMachineLearningResearch,pages1597–1607.PMLR,2020

    work page 2020

  7. [7]

    Exploring simple siamese repre- sentationlearning

    Xinlei Chen and Kaiming He. Exploring simple siamese repre- sentationlearning. InProceedingsoftheIEEE/CVFConference on Computer Vision and Pattern Recognition (CVPR), pages 15750–15758, 2021. doi: 10.1109/CVPR46437.2021.01549

    arXiv 2021

  8. [8]

    An empirical study of training self-supervised vision transformers

    Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9620–9629, 2021. doi: 10.1109/ICCV48922. 2021.00950

    work page arXiv 2021

Show all 51 references
  1. [9]

    Time-series representation learning via temporal and contextual contrasting

    EmadeldeenEldele,MohamedRagab,ZhenghuaChen,MinWu, Chee Keong Kwoh, Xiaoli Li, and Cuntai Guan. Time-series representation learning via temporal and contextual contrasting. InProceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), pages 2...

    2021 doi

  2. [10]

    CLAP: Learning audio concepts from naturallanguagesupervision

    Benjamin Elizalde, Soham Deshmukh, Mahmoud Al Ismail, and Huaming Wang. CLAP: Learning audio concepts from naturallanguagesupervision. InIEEEInternationalConference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5, 2023. doi: 10.1109/ICASSP49357.2023.10095889

    2023

  3. [11]

    Joint embedding go temporal

    Sofiane Ennadir, Siavash Golkar, and Leopoldo Sarra. Joint embedding go temporal. InNeurIPS Workshop on Time Series in the Age of Large Models, 2024. URL https://openreview.net/ forum?id=FIdbozebmy

    2024

  4. [12]

    Moment:Afamilyofopentime- series foundation models.arXiv preprint arXiv:2402.03885, 2024

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, ShuoLi,andArturDubrawski. Moment:Afamilyofopentime- series foundation models.arXiv preprint arXiv:2402.03885, 2024

    2024 arXiv

  5. [13]

    Nathanson, Alistair Johnson, BenjaminMoody,ChrystinneFernandes,NathanielGreenbaum, Jonathan W

    Brian Gow, Tom Pollard, Larry A. Nathanson, Alistair Johnson, BenjaminMoody,ChrystinneFernandes,NathanielGreenbaum, Jonathan W. Waks, Parastou Eslami, Tanner Carbonati, Ashish Chaudhari,ElizabethHerbst,DanaMoukheiber,SethBerkowitz, Roger Mark, and Steven Horng. MIMIC-IV-ECG: D...

    2023 doi

  6. [14]

    Richemond, Elena Buchatskaya, Carl Doersch, BernardoAvilaPires,ZhaohanDanielGuo,MohammadGhesh- laghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tal- lec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, BernardoAvilaPires,ZhaohanDanielGuo,MohammadGhesh- laghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A ...

    2020

  7. [15]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979–15988, 2022. doi: 10.1109/CVPR52688.2022.01553

    2022

  8. [16]

    Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichten- hofer.Maskedautoencodersthatlisten.InAdvancesinNeuralIn- formation Processing Systems, volume 35, pages 28708–28720, 2022

    2022

  9. [17]

    Reading your heart: Learning ECG words and sentencesviapre-trainingECGlanguagemodel.InInternational Conference on Learning Representations (ICLR), 2025

    Jiarui Jin, Haoyu Wang, Hongyan Li, Jun Li, Jiahui Pan, and Shenda Hong. Reading your heart: Learning ECG words and sentencesviapre-trainingECGlanguagemodel.InInternational Conference on Learning Representations (ICLR), 2025. URL https://openreview.net/forum?id=6Hz1Ko087B

    2025

  10. [18]

    Clifford, and Reza Sameni

    Arsalan Kazemnejad, Sajjad Karimi, Peiman Gordany, Gari D. Clifford, and Reza Sameni. An open-access simultaneous electrocardiogram and phonocardiogram database.Physiolog- ical Measurement, 45(5):055005, 2024. doi: 10.1088/1361- 6579/ad43af

    2024 doi

  11. [19]

    Dani Kiyasseh, Tingting Zhu, and David A. Clifton. CLOCS: Contrastive learning of cardiac signals across space, time, and patients. InProceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pages 5606–5615. P...

    2021

  12. [20]

    AleksandarLazović,PredragTadić,NatalijaÐorđević,Vladimir Atanasoski, Masa Tiosavljevic, Marija Ivanovic, Ljupco Hadzievski, Arsen Ristic, Vladan Vukcevic, and Jovana Petro- vic. SensSmartTech database of cardiovascular signals syn- chronously recorded by an electrocardiograph,...

    2024 doi

  13. [21]

    VitalDB, a high- fidelity multi-parameter vital signs database in surgical patients

    Hyung-Chul Lee, Yoonsang Park, Soo Bin Yoon, Seong Mi Yang, Dongnyeok Park, and Chul-Woo Jung. VitalDB, a high- fidelity multi-parameter vital signs database in surgical patients. Scientific Data, 9(1):279, 2022. doi: 10.1038/s41597-022- 01411-5

    2022 doi

  14. [22]

    anyECG-chat:AgeneralistECG-MLLMfor flexibleECGinputandmulti-taskunderstanding.InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 597–605, 2026

    Haitao Li, Ziyu Li, Yiheng Mao, Ziyi Liu, Zhoujian Sun, and ZhengxingHuang. anyECG-chat:AgeneralistECG-MLLMfor flexibleECGinputandmulti-taskunderstanding.InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 597–605, 2026. doi: 10.1609/aaai.v40i1.37024

    2026 doi

  15. [23]

    Clifford, M

    JunLi,AaronD.Aguirre,ValderyMouraJunior,JiaruiJin,Che Liu, Lanhai Zhong, Chenxi Sun, Gari D. Clifford, M. Brandon Westover, and Shenda Hong. An electrocardiogram foundation model built on over 10 million recordings.NEJM AI, 2(7): AIoa2401033, 2025. doi: 10.1056/AIoa2401033

    2025 doi

  16. [24]

    Wearablesensorsanddevicesforreal-timecardiovascular disease monitoring.Cell reports physical science, 2(8), 2021

    Jian Lin, Rumin Fu, Xinxiang Zhong, Peng Yu, Guoxin Tan, Wei Li, Huan Zhang, Yangfan Li, Lei Zhou, and Chengyun Ning. Wearablesensorsanddevicesforreal-timecardiovascular disease monitoring.Cell reports physical science, 2(8), 2021

    2021

  17. [25]

    Zero-shot ECG classification with multimodal learning and test-time clinical knowledge enhance- ment

    Che Liu, Zhongwei Wan, Cheng Ouyang, Anand Shah, Wenjia Bai, and Rossella Arcucci. Zero-shot ECG classification with multimodal learning and test-time clinical knowledge enhance- ment. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofProceed...

    2024

  18. [26]

    Chorro, Francisco Castells, Jose Millet Roig, Ikaro Silva, Alistair E

    Chengyu Liu, David Springer, Qiao Li, Benjamin Moody, Ri- cardo Abad Juan, Francisco J. Chorro, Francisco Castells, Jose Millet Roig, Ikaro Silva, Alistair E. W. Johnson, Zee- shan Syed, Samuel E. Schmidt, Chrysa D. Papadaniil, Leontios Hadjileontiadis,HoseinNaseri,AliMoukadem...

    2016

  19. [27]

    Feifei Liu, Chengyu Liu, Lina Zhao, Xiangyu Zhang, Xiaoling Wu,XiaoyanXu,YulinLiu,CaiyunMa,ShoushuiWei,Zhiqiang He, Jianqing Li, and Eddie Ng Yin. An open access database for evaluating the algorithms of electrocardiogram rhythm and morphologyabnormalitydetection.JournalofMedi...

    2018

  20. [28]

    ECG-FM: An open electrocardiogram foundation model.JAMIA Open, 8(5):ooaf122, 2025

    Kaden McKeen, Sameer Masood, Augustin Toma, Barry Ru- bin, and Bo Wang. ECG-FM: An open electrocardiogram foundation model.JAMIA Open, 8(5):ooaf122, 2025. doi: 10.1093/jamiaopen/ooaf122

    2025 doi

  21. [29]

    Aston, Pe- ter H

    Mohammad Moulaeifard, Marie Kutscher, Philip J. Aston, Pe- ter H. Charlton, and Nils Strodthoff. MIMIC-III-Ext-PPG, a PPG-based benchmark dataset for cardiovascular and res- piratory signal analysis.Scientific Data, 13:668, 2026. doi: 10.1038/s41597-026-07335-8

    2026 doi

  22. [30]

    Guiding masked representation learning to capture spatio- temporal relationship of electrocardiogram

    Yeongyeon Na, Minje Park, Yunwon Tae, and Sunghoon Joo. Guiding masked representation learning to capture spatio- temporal relationship of electrocardiogram. InInternational Conference on Learning Representations (ICLR), 2024. URL https://openreview.net/forum?id=WcOohbsF4H

    2024

  23. [31]

    AnyPPG: An ECG-guided PPG foundation model trained on over 100,000 hours of recordings for holistic health profiling.arXiv preprint arXiv:2511.01747, 2025

    Guangkun Nie, Gongzheng Tang, Yujie Xiao, Jun Li, Shun Huang, Deyun Zhang, Qinghao Zhao, and Shenda Hong. AnyPPG: An ECG-guided PPG foundation model trained on over 100,000 hours of recordings for holistic health profiling.arXiv preprint arXiv:2511.01747, 2025. doi: 10.48550/a...

    2025 doi arXiv

  24. [32]

    Clifford, and Miguel Coimbra

    Jorge Oliveira, Francesco Renna, Paulo Costa, Marcelo Nogueira, Cristina Oliveira, Andoni Elola, Carlos Ferreira, Alipio Jorge, Ali Bahrami Rad, Reza Sameni, Gari D. Clifford, and Miguel Coimbra. The CirCor DigiScope phonocardiogram dataset.PhysioNet, 2022. doi: 10.13026/tshs-...

    2022 doi

  25. [33]

    Boosting masked ECG-text auto-encoders as discriminative learners

    Hung Manh Pham, Aaqib Saeed, and Dong Ma. Boosting masked ECG-text auto-encoders as discriminative learners. InProceedings of the 42nd International Conference on Ma- chine Learning, 2025. URL https://openreview.net/forum?id= mM65b81LdM

    2025

  26. [34]

    PulseLM: A foundation dataset and benchmark for PPG-text learning

    HungManhPham,JinyangWu,XiaoMa,YimingZhang,Yixin Xu, Aaqib Saeed, Bin Zhu, Zhou Pan, and Dong Ma. PulseLM: A foundation dataset and benchmark for PPG-text learning. arXiv preprint arXiv:2603.03331, 2026. doi: 10.48550/arXiv. 2603.03331

    2026 doi arXiv

  27. [35]

    PaPaGei: Open foundation models for optical physiological signals

    ArvindPillai,DimitrisSpathis,FahimKawsar,andMohammad Malekzadeh. PaPaGei: Open foundation models for optical physiological signals. InInternational Conference on Learning Representations (ICLR), 2025. URL https://openreview.net/ forum?id=kYwTmlq6Vn

    2025

  28. [36]

    Reyna, Yashar Kiarashi, Andoni Elola, Jorge Oliveira, Francesco Renna, Annie Gu, Erick A

    Matthew A. Reyna, Yashar Kiarashi, Andoni Elola, Jorge Oliveira, Francesco Renna, Annie Gu, Erick A. Perez Al- day, Nadi Sadr, Ashish Sharma, Jacques Kpodonu, Sandra Mattos, Miguel T. Coimbra, Reza Sameni, Ali Bahrami Rad, and Gari D. Clifford. Heart murmur detection from phon...

    2022 doi

  29. [37]

    Xu, Wanting Mao, Sameer Neu- pane, James M

    Mithun Saha, Maxwell A. Xu, Wanting Mao, Sameer Neu- pane, James M. Rehg, and Santosh Kumar. Pulse-PPG: An open-source field-trained PPG foundation model for wearable applications across lab and field settings.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquit...

    2025 doi

  30. [38]

    Intro- ducing wesad, a multimodal dataset for wearable stress and affect detection

    Philipp Schmidt, Attila Reiss, Robert Dürichen, et al. Intro- ducing wesad, a multimodal dataset for wearable stress and affect detection. InProceedings of the 20th ACM International Conference on Multimodal Interaction (ICMI), pages 400–408,

  31. [39]

    Time-moe: Billion-scale time series foundation models with mixture of experts

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts. InInterna- tional conference on learning representations, volume 2025, pages 34635–34667, 2025

    2025

  32. [40]

    Lunze, Wojciech Samek, and To- bias Schaeffter

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Di- eter Kreiseler, Fatima I. Lunze, Wojciech Samek, and To- bias Schaeffter. PTB-XL, a large publicly available electro- cardiography dataset.Scientific Data, 7(1):154, 2020. doi: 10.1038/s41597-020-0495-6. 9

    2020 doi

  33. [41]

    Adversarial spatiotemporal con- trastive learning for electrocardiogram signals.IEEE Trans- actions on Neural Networks and Learning Systems, 35(10): 13845–13859, 2024

    Ning Wang, Panpan Feng, Zhaoyang Ge, Yanjie Zhou, Bing Zhou, and Zongmin Wang. Adversarial spatiotemporal con- trastive learning for electrocardiogram signals.IEEE Trans- actions on Neural Networks and Learning Systems, 35(10): 13845–13859, 2024. doi: 10.1109/TNNLS.2023.3272153

    2024

  34. [42]

    Stetholm: Audio language model for cardiopulmonary analysisacrossclinicaltasks.TransactionsonMachineLearning Research, 2026

    Yishan Wang, Tsai-Ning Wang, Mathias Funk, and Aaqib Saeed. Stetholm: Audio language model for cardiopulmonary analysisacrossclinicaltasks.TransactionsonMachineLearning Research, 2026. URL https://arxiv.org/abs/2603.00355

    2026

  35. [43]

    Cost:Contrastivelearningofdisentangledseasonal- trend representations for time series forecasting.arXiv preprint arXiv:2202.01575, 2022

    Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and StevenHoi. Cost:Contrastivelearningofdisentangledseasonal- trend representations for time series forecasting.arXiv preprint arXiv:2202.01575, 2022

    2022 arXiv

  36. [44]

    ECGsemanticintegrator (ESI):AfoundationECGmodelpretrainedwithLLM-enhanced cardiologicaltext.TransactionsonMachineLearningResearch,

    HanYu,PeikunGuo,andAkaneSano. ECGsemanticintegrator (ESI):AfoundationECGmodelpretrainedwithLLM-enhanced cardiologicaltext.TransactionsonMachineLearningResearch,

  37. [45]

    Ts2vec:Towards universal representation of time series

    Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, CongruiHuang,YunhaiTong,andBixiongXu. Ts2vec:Towards universal representation of time series. InProceedings of the AAAI conference on artificial intelligence, volume 36, pages 8980–8987, 2022

    2022

  38. [46]

    Barlow twins: Self-supervised learning via redundancy reduction

    Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. InProceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pages 12310–12320....

    2021

  39. [47]

    Self-supervised time series representation learning via cross reconstruction transformer.IEEE Transactions on Neural Net- works and Learning Systems, 35(11):16129–16138, 2024

    Wenrui Zhang, Ling Yang, Shijia Geng, and Shenda Hong. Self-supervised time series representation learning via cross reconstruction transformer.IEEE Transactions on Neural Net- works and Learning Systems, 35(11):16129–16138, 2024. doi: 10.1109/TNNLS.2023.3292066

    2024

  40. [48]

    A 12-lead electro- cardiogram database for arrhythmia research covering more than 10,000 patients.Scientific Data, 7(1):48, 2020

    Jianwei Zheng, Jianming Zhang, Sidy Danioko, Hai Yao, Hangyuan Guo, and Cyril Rakovski. A 12-lead electro- cardiogram database for arrhythmia research covering more than 10,000 patients.Scientific Data, 7(1):48, 2020. doi: 10.1038/s41597-020-0386-x. 10 A Appendix A.1 Data and ...

    2020 doi

  41. [2018]

    doi: 10.1145/3242969.3242985

  42. [2024]

    URL https://openreview.net/forum? id=giEbq8Khcf

    ISSN 2835-8856. URL https://openreview.net/forum? id=giEbq8Khcf

  43. [9900]

    doi: 10.1016/j.cmpbup.2026.100237

    2026

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.