Pith. sign in

REVIEW 4 major objections 4 minor 6 cited by

AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims PPG signals—trained with ECG as a teacher—can hint at 137 diseases across the body, not just heart conditions.

desk verdict ECG-guided pretraining is a real contribution, but the phenome-wide disease claim is confounded and the abstract numbers are inconsistent. read the letter →

arxiv 2511.01747 v4 pith:X3I2DME5 submitted 2025-11-03 eess.SP

Guangkun Nie , Xiaocheng Fang , Gongzheng Tang , Yujie Xiao , Jun Li , Bo Liu , Hongyan Li , Shenda Hong This is my paper
classification eess.SP
keywords photoplethysmographyfoundationmodelECG-guidedpretrainingcontrastivelearningphenome-wideassociationstudyICD-10multi-organhealthprofilingwearablemonitoring
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 sets out to answer whether photoplethysmography (PPG), the optical pulse signal in wearable heart-rate monitors, can support health profiling beyond cardiovascular tasks. The authors build AnyPPG, a foundation model pretrained on over 100,000 hours of synchronized PPG-ECG recordings using a contrastive objective that aligns the two signals in a shared embedding space. The model beats previous approaches on 11 conventional physiological tasks, and in a phenome-wide screen of roughly 16,000 held-out patients it reaches AUC ≥ 0.7 for 137 of 719 analyzable three-digit ICD-10 codes, including Parkinson's disease, chronic kidney disease, osteoporosis, and pregnancy complications. The paper takes this as evidence that PPG encodes multi-organ health information, which would make cheap, ubiquitous wearables a plausible screening modality.

What carries the argument

The central mechanism is a CLIP-style contrastive alignment between two structurally identical one-dimensional convolutional encoders. Synchronized 10-second PPG and ECG segments are embedded, projected into a shared 256-dimensional space, and matched via a symmetric contrastive loss with a learnable temperature. The ECG branch acts as a physiological teacher: because the two signals are temporally aligned, the PPG encoder is forced to recover cardiac electrical timing and rhythm information from the optical waveform alone. At downstream inference the ECG branch is discarded, so the PPG encoder carries the distilled cardio-systemic knowledge.

What would settle it

A matched external cohort (balanced for age, sex, and acuity) in which a model trained only on demographics and standard vital signs—without the PPG waveform—matches or beats AnyPPG's AUCs for non-cardiovascular diseases would falsify the claim that the optical signal itself carries the diagnostic information.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that cross-modal alignment—training the PPG encoder to match ECG embeddings from the same 10-second window through a bidirectional contrastive objective—yields PPG representations that transfer to a strikingly broad set of disease phenotypes. The strongest evidence is the phenome-wide analysis: after fine-tuning on disease-annotated subjects, the model achieves AUC ≥ 0.7 for 137 three-digit ICD-10 codes, and 82 remain after removing non-specific labels; these span 14 chapters, from neoplasms and eye diseases to pregnancy and musculoskeletal disorders. The authors interpret this as showing that peripheral hemodynamics, as captured by PPG, ref

Load-bearing premise

The phenome-wide disease screen assumes that ICD billing labels from a single emergency-department dataset are accurate enough, and that the held-out-subject split removes confounders such as age, illness acuity, and comorbidity; if those confounders drive the AUCs rather than disease-specific PPG information, the holistic-health conclusion collapses.

Editorial extensions

If this is right

  • If PPG really carries multi-organ information, wearable devices that already record PPG could be repurposed for screening of neurological, renal, respiratory, and other non-cardiovascular conditions.
  • The ECG-guided pretraining recipe—align a cheap peripheral signal to a gold-standard signal—is a transferable strategy for building foundation models for other biosignals.
  • The large gains on heart-rate estimation and atrial fibrillation detection indicate that cross-modal supervision adds information beyond single-modality pretraining.
  • After excluding non-specific ICD labels, 82 diseases still clear AUC 0.7, so the finding is not purely an artifact of vague billing codes.
  • The held-out-subject design means the phenome-wide estimates are not inflated by the model having seen the same patients during pretraining.

Reading between the lines

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

  • An alternative reading the paper acknowledges indirectly: many high-AUC diseases (dementia, CKD, osteoporosis, heart failure) are markers of older, frailer patients, so the AUCs may partly reflect a systemic 'illness load' rather than disease-specific waveform signatures; a matched-control or age-stratified analysis would separate these.
  • Because the ECG branch is discarded at inference, the approach suggests a general principle: distill information from an invasive or expensive sensor into a cheap one during pretraining, then deploy the cheap sensor alone.
  • A testable extension is longitudinal risk prediction: rather than classifying a current ICD code, the same representations could be used to predict future disease onset from wearable PPG, which would convert the screening claim into a prevention claim.
  • The single emergency-department dataset limits generalizability; validating on outpatient or wearable-collected PPG from diverse populations is the natural next step, and would likely lower the optimistic AUCs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. AnyPPG is a PPG foundation model pretrained with ECG-guided contrastive learning on 109,909 hours of synchronized PPG-ECG data from five public datasets (58,796 subjects). The authors evaluate it on eleven conventional physiological tasks (heart rate, blood pressure, stress/affect recognition, AF detection, signal quality) against two PaPaGei baselines, and on a phenome-wide disease diagnosis task using 719 ICD-10 three-digit codes from the MC-MED emergency department dataset. They report state-of-the-art results on the conventional tasks and AUC ≥ 0.70 for 137 ICD-10 codes (82 after excluding non-specific codes), including many non-cardiovascular conditions, and conclude that PPG supports holistic, multi-organ health profiling.

Significance. The conventional-task evidence is credible and a useful contribution: six independent datasets, subject-level splits, nested five-fold CV, and public code/weights. If the phenome-wide claim were robust, the paper would substantially broaden the perceived utility of PPG beyond cardiovascular monitoring. However, the central claim currently rests on a single-dataset, confounder-unadjusted, segment-level analysis with no demographic-only baseline and no multiple-testing control. The strengths are the scale of pretraining, the ECG-guided design, and the release of code and pretrained weights; the phenome-wide conclusion is not yet secured.

major comments (4)
  1. [Results, 'AnyPPG reveals the potential of PPG for comprehensive multi-organ disease diagnosis'; Methods, 'Model evaluati] The central claim that PPG supports holistic multi-organ health profiling rests entirely on the MC-MED phenome-wide screen. The reported AUCs are unadjusted for age, sex, comorbidity, or acuity, and no demographic-only baseline is presented. Many of the top non-cardiovascular codes (e.g., F03 unspecified dementia 0.80, M81 osteoporosis 0.79, G20 Parkinson 0.78, H25 cataract 0.76, N18 CKD 0.74 in Table A1) are strongly associated with age and multimorbidity. In a single emergency department, the model may separate 'old and frail' from 'young and healthy' rather than detecting disease-specific PPG signatures. Please add a demographic-only baseline (age/sex, and ideally comorbidity/acuity) and report adjusted or stratified AUCs.
  2. [Table A1 and Methods ('20 PPG segments randomly sampled from each hospitalization record')] The phenome-wide AUCs are computed at the segment level: 359,900 segments from 15,759 subjects, with 20 segments sampled per record. No confidence intervals, no clustering adjustment, no subject-level aggregation, and no multiple-testing control across 719 tested ICD-10 codes are reported. Because segments from the same patient are highly correlated, the AUC estimates are likely overconfident. Please report cluster-robust or subject-level AUCs with confidence intervals and apply an FDR-controlling procedure when enumerating the 137 (or 82) 'significant' codes.
  3. [Abstract vs. Results vs. Discussion] The numbers supporting the headline claim are internally inconsistent. The Abstract states that AnyPPG achieves AUC ≥ 0.70 for '307 phenotypes across 16 distinct phecode chapters, including 230 non-circulatory conditions,' while the Results report '137 diseases' among 719 ICD-10 codes (82 after excluding non-specific labels) across ICD Chapters I–XV (15 chapters), and the Discussion says '133 diseases.' These must be reconciled, and the 'phecode' terminology should be aligned with the ICD-10 analysis actually performed.
  4. [Discussion, limitations paragraph] The phenome-wide analysis uses a single dataset (MC-MED, Stanford ED) and the Discussion explicitly concedes that it 'lacks external, multi-center validation.' Because the manuscript's central conclusion is about the breadth and generalizability of PPG-based health profiling, this is not just a routine limitation: without at least one independent cohort, or a clear demonstration that the results are not an artifact of one hospital's population and recording system, the holistic-health claim is not yet established. Please provide external validation or substantially soften the conclusion.
minor comments (4)
  1. [Results, Figure 1/Table 3/Table 4] The 'state-of-the-art' claim is based on a comparison with only PaPaGei-S and PaPaGei-P; GPT-PPG and PulsePPG are cited but not benchmarked. Please qualify the claim or include these baselines.
  2. [Abstract and Figure 2] Please define 'phenotype' and 'phecode chapter' precisely; the main text uses ICD-10 three-digit codes and ICD chapters, so the abstract's phecode language is confusing.
  3. [Discussion, third paragraph] '133 diseases' should be '137 diseases' (or the corrected number after reconciliation with the Abstract).
  4. [Figure 2 and Table A1] The top-50 figure says it excludes non-specific diagnostic codes, but Table A1 includes many 'other/unspecified' codes. State the exact exclusion rule and apply it consistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the ECG-guided pretraining, held-out-subject disease screen, and downstream benchmarks are empirical evaluations rather than reductions to their own inputs.

full rationale

I walked the claimed derivation chain: (1) AnyPPG is pretrained with a symmetric InfoNCE loss that aligns synchronized PPG and ECG embeddings; this objective contains no ICD-10 labels and no downstream task labels, so the later disease AUCs cannot be tautological outputs of the pretraining loss. (2) The multi-organ disease analysis is supervised by external ICD-10 codes and evaluated on subjects explicitly held out from pretraining: 'evaluation strictly relied on subjects and recordings that were entirely unseen during pre-training (including all train, validation, and test partitions) to ensure subject-level independence.' Thus the 137 ICD codes with AUC>0.7 are empirical correlations, not definitions. (3) The conventional physiological evaluations use six independent public datasets and are compared against baseline models, so those results are externally benchmarked rather than forced by construction. (4) The paper's self-citations (e.g., Net1D architecture, earlier PPG/sleep work) are used for architecture or background context; no load-bearing claim is justified solely by a self-citation, and no uniqueness theorem is imported from the authors' prior work. (5) The Discussion explicitly acknowledges that the MC-MED phenome-wide analysis 'lacks external, multi-center validation' and that diagnostic labels 'may contain inherent variability.' These are external-validity and label-quality limitations, and the absence of demographic/comorbidity controls is a confounding-control concern, but none of these constitute a circular derivation. No specific equation or parameter is fitted to the target and then renamed as a prediction. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities or forces are postulated. The central free parameters are standard training hyperparameters plus hand-chosen analysis thresholds (prevalence cutoff, segment count). The load-bearing assumptions are domain-level: ECG transferability, MC-MED label quality, and adequacy of preprocessing.

free parameters (3)
  • Contrastive temperature tau = learned (initialized to 0.07)
    Learnable temperature in the InfoNCE loss; standard in CLIP-style training, but its fitted value influences embedding sharpness and retrieval/downstream behavior.
  • Disease-code prevalence threshold = >= 100 positive test samples
    ICD-10 codes with fewer than 100 positive samples in the MC-MED test set were excluded, reducing 1,014 codes to 719. This is a hand-chosen robustness threshold that shapes the reported disease landscape.
  • PPG segments per hospitalization record = 20
    For MC-MED disease evaluation, 20 PPG segments were randomly sampled from each hospitalization record. This sampling choice affects variance and AUC estimates but is not varied in the paper.
assumptions (4)
  • domain assumption Synchronized 10-second PPG and ECG windows share sufficient physiological information that contrastive alignment transfers to PPG-only downstream tasks.
    This is the core mechanistic premise for ECG-guided pretraining. It is supported empirically by retrieval metrics (R@1=0.736) and by conventional-task gains, but it is not proven that the transferred features are disease-specific.
  • domain assumption MC-MED ICD-10 diagnoses are an adequate ground-truth label for measuring the discriminative value of PPG.
    All phenome-wide AUCs depend on the accuracy of emergency-department ICD billing labels. The authors themselves note 'inherent variability' in these labels and the absence of external multi-center validation.
  • standard math Standard InfoNCE/contrastive learning and the Net1D/ResNet-style architecture are valid modeling tools.
    The paper does not prove properties of InfoNCE or the convolutional encoder; it treats them as background ML machinery.
  • domain assumption The preprocessing choices (0.5-8 Hz PPG band-pass, 0.5-40 Hz ECG band-pass, 50 Hz notch, 125 Hz resampling) preserve disease-relevant information.
    Filtering follows prior references; if disease-relevant slow or high-frequency PPG components were removed, downstream AUCs could be affected.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling." pith.science (2026). https://pith.science/paper/X3I2DME5

@misc{pith2026251101747,
  author       = {Pith},
  title        = {Pith review of: AnyPPG: An ECG-Guided PPG Foundation Model Trained on Over 100,000 Hours of Recordings for Holistic Health Profiling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X3I2DME5}},
  note         = {Machine review of arXiv:2511.01747}
}
abstract

Photoplethysmography (PPG) is widely used as a non-invasive and accessible modality for continuous health monitoring. However, despite being a peripheral hemodynamic signal intrinsically coupled with systemic circulation, existing research has largely confined its scope to a narrow range of cardiovascular tasks, leaving a fundamental question underexplored: to what extent can PPG support holistic health profiling beyond traditional cardiovascular applications? To answer this question, we present AnyPPG, a foundation model-based framework designed to reveal the broader health-profiling potential of PPG. To ensure reliable performance for this investigation, AnyPPG is pretrained with ECG guidance on the most diverse PPG corpus with synchronized ECG to date, comprising over 100,000 hours of recordings from six large-scale data sources. This pretraining yields robust and physiologically grounded PPG representations that provide a reliable basis for subsequent analysis. Building upon this pretrained model, we conduct a systematic investigation into the association between PPG and holistic health through, to our knowledge, the first PPG-based phenome-wide disease detection study, spanning 1,468 disease phenotypes in more than 15,000 subjects. Our evaluation demonstrates the effectiveness of AnyPPG: across eight clinical and wearable datasets covering 15 downstream tasks, it achieves the best performance in 13 tasks. More importantly, in the phenome-wide analysis, AnyPPG exhibits meaningful discriminative capability (AUC $\ge$ 0.70) for 307 phenotypes across 16 distinct phecode chapters, including 230 non-circulatory conditions such as dementia and chronic kidney disease, many of which have rarely been explored using PPG. Collectively, these findings indicate that easily acquired PPG signals encode rich health-related information extending well beyond conventional cardiovascular assessment.

Figures

Figures reproduced from arXiv: 2511.01747 by the authors.

Figure 1
Figure 1. Performance comparison of AnyPPG and baseline models across downstream tasks [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Diagnostic performance of AnyPPG across ICD-10 disease categories. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 6 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. BCG-FM: A Foundation Model for Ambient Cardiac Health Sensing

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    BCG-FM, the first foundation model for ambient BCG, achieves 3.26-year MAE on biological age estimation and discriminates 15 health conditions using frozen embeddings from participant-level contrastive pretraining on ...

  2. Physiology-Aware Masked Cross-Modal Reconstruction for Biosignal Representation Learning

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    xMAE pretrains biosignal representations via masked cross-modal reconstruction of temporally ordered signals like ECG and PPG, outperforming baselines on 15 of 19 downstream tasks including cardiovascular prediction a...

  3. ImputeECG: Deep Learning Reconstruction of Complete 12-Lead Electrocardiograms from Incomplete Recordings for Cardiac Assessment

    cs.LG 2026-07 conditional novelty 6.0 of 10

    A mask-conditioned 1D Transformer autoencoder completes incomplete 12-lead 10-s ECGs, cutting missing-region MAE 42–52% and restoring multi-label diagnostic AUROC near complete-ECG levels on PTB-XL, CPSC2018, and a 43...

  4. WavesFM: Hierarchical Representation Learning for Longitudinal Wearable Sensor Waveforms

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    WavesFM uses hierarchical SSL to pretrain a segment encoder on short waveforms followed by a temporal encoder on multi-day sequences, outperforming prior methods on 58 tasks after training on over 12 million hours of ...

  5. PRISM-CTG: A Foundation Model for Cardiotocography Analysis with Multi-View SSL

    cs.LG 2026-04 unverdicted novelty 6.0 of 10

    PRISM-CTG is the first large-scale foundation model for cardiotocography that uses multi-view self-supervised learning on unlabeled data to learn transferable representations, outperforming baselines on seven downstre...

  6. EchoBridge: Long-Tail-Aware ECG-Echocardiography Text Alignment for Echocardiography-Derived Cardiac Findings

    cs.LG 2026-07 conditional novelty 5.5 of 10

    EchoBridge’s shared–private ECG–echo-text alignment plus frequency-adaptive prototypes beats strong baselines on classifier-free and cross-center frozen probing, including several low-prevalence valvular findings.

Reference graph

Works this paper leans on

49 extracted references · 3 linked inside Pith · cited by 6 Pith papers

  1. [1]

    S., Ginsburg, G

    Spatz, E. S., Ginsburg, G. S., Rumsfeld, J. S. & Turakhia, M. P . Wearable digital health 395 technologies for monitoring in cardiovascular medicine.New England Journal of Medicine 396 390, 346–356 (2024). 397

    2024

  2. [2]

    J.et al.Wearable technology and the cardiovascular system: the future of 398 patient assessment.The Lancet Digital Health5, e467–e476 (2023).399

    Williams, G. J.et al.Wearable technology and the cardiovascular system: the future of 398 patient assessment.The Lancet Digital Health5, e467–e476 (2023).399

    2023

  3. [3]

    Pedroso, A. F . & Khera, R. Leveraging ai-enhanced digital health with consumer devices for 400 scalable cardiovascular screening, prediction, and monitoring.npj Cardiovascular Health2, 401 34 (2025). 402

    2025

  4. [4]

    Bayoumy, K.et al.Smart wearable devices in cardiovascular care: where we are and how to 403 move forward.Nature Reviews Cardiology18, 581–599 (2021).404

    2021

  5. [5]

    Nie, G.et al.A review of deep learning methods for photoplethysmography data.arXiv 405 preprint arXiv:2401.12783(2024). 406

    arXiv 2024

  6. [6]

    S., Kim, S.-S

    Park, J., Seok, H. S., Kim, S.-S. & Shin, H. Photoplethysmogram analysis and applications: 407 an integrative review.Frontiers in physiology12, 808451 (2022).408

    2022

  7. [7]

    & Kumar, M

    Pankaj, Kumar, A., Komaragiri, R. & Kumar, M. A review on computation methods used in 409 photoplethysmography signal analysis for heart rate estimation.Archives of Computational 410 Methods in Engineering29, 921–940 (2022). 411

    2022

  8. [8]

    Elgendi, M.et al.The use of photoplethysmography for assessing hypertension.NPJ digital 412 medicine2, 60 (2019). 413

    2019

Show all 49 references
  1. [9]

    Pereira, T.et al.Photoplethysmography based atrial fibrillation detection: a review.NPJ 414 digital medicine3, 3 (2020). 415

    2020

  2. [10]

    & Hong, S

    Nie, G., Zhao, Q., Tang, G., Li, Y . & Hong, S. Artificial intelligence-derived vascular age from 416 photoplethysmography: A novel digital biomarker for cardiovascular health.arXiv preprint 417 arXiv:2502.12990(2025). 418

    2025

  3. [11]

    He, Y .et al.Foundation model for advancing healthcare: Challenges, opportunities and 419 future directions.IEEE Reviews in Biomedical Engineering(2024).420

    2024

  4. [12]

    Moor, M.et al.Foundation models for generalist medical artificial intelligence.Nature616, 421 259–265 (2023). 422

    2023

  5. [13]

    Xiang, J.et al.A vision–language foundation model for precision oncology.Nature638, 423 769–778 (2025). 424

    2025

  6. [14]

    Vorontsov, E.et al.A foundation model for clinical-grade computational pathology and rare 425 cancers detection.Nature medicine30, 2924–2935 (2024).426

    2024

  7. [15]

    Y .et al.A multimodal generative ai copilot for human pathology.Nature634, 466–473 427 (2024)

    Lu, M. Y .et al.A multimodal generative ai copilot for human pathology.Nature634, 466–473 427 (2024). 428

    2024

  8. [16]

    & Ouyang, D

    Christensen, M., Vukadinovic, M., Yuan, N. & Ouyang, D. Vision–language foundation model 429 for echocardiogram interpretation.Nature Medicine30, 1481–1488 (2024).430 16

    2024

  9. [17]

    InInternational Conference on Machine Learning, 48019–48037 432 (PMLR, 2024)

    Thapa, R.et al.Sleepfm: Multi-modal representation learning for sleep across brain activity, 431 ecg and respiratory signals. InInternational Conference on Machine Learning, 48019–48037 432 (PMLR, 2024). 433

    2024

  10. [18]

    Nie, G.et al.A low-burden sleep foundation model built on respiratory and heartbeat signals 434 from 780,000+ hours of multi-ethnic sleep recordings.medRxiv2025–09 (2025).435

    2025

  11. [19]

    Li, J.et al.An electrocardiogram foundation model built on over 10 million recordings.NEJM 436 AI2, AIoa2401033 (2025). 437

    2025

  12. [20]

    Zhang, S.et al.Ecgfm: A foundation model for ecg analysis trained on a multi-center 438 million-ecg dataset.Information Fusion103363 (2025). 439

    2025

  13. [21]

    Huang, S.et al.Combining ecg foundation model and xgboost to predict in-hospital malignant 440 ventricular arrhythmias in ami patients.arXiv preprint arXiv:2510.17172(2025).441

    2025

  14. [22]

    Xiao, Y .et al.Anyecg-lab: An exploration study of fine-tuning an ecg foundation model to 442 estimate laboratory values from single-lead ecg signals.arXiv preprint arXiv:2510.22301 443 (2025). 444

    2025

  15. [23]

    & Malekzadeh, M

    Pillai, A., Spathis, D., Kawsar, F . & Malekzadeh, M. Papagei: Open foundation models 445 for optical physiological signals. InThe Thirteenth International Conference on Learning 446 Representations. 447

  16. [24]

    448 Physiological Measurement46, 055004 (2025)

    Chen, Z.et al.Gpt-ppg: a gpt-based foundation model for photoplethysmography signals. 448 Physiological Measurement46, 055004 (2025). 449

    2025

  17. [25]

    Saha, M.et al.Pulse-ppg: An open-source field-trained ppg foundation model for wearable 450 applications across lab and field settings.Proceedings of the ACM on Interactive, Mobile, 451 Wearable and Ubiquitous Technologies9, 1–35 (2025). 452

    2025

  18. [26]

    In 453 International conference on machine learning, 8748–8763 (PmLR, 2021).454

    Radford, A.et al.Learning transferable visual models from natural language supervision. In 453 International conference on machine learning, 8748–8763 (PmLR, 2021).454

    2021

  19. [27]

    Yu, J.et al.Coca: Contrastive captioners are image-text foundation models.arXiv preprint 455 arXiv:2205.01917(2022). 456

    2022 arXiv

  20. [28]

    Wang, F ., Xu, J. & Yu, L. From token to rhythm: A multi-scale approach for ecg-language 457 pretraining. InForty-second International Conference on Machine Learning.458

  21. [29]

    T., Rajpurkar, P

    Kansal, A., Chen, E., Jin, B. T., Rajpurkar, P . & Kim, D. A. Mc-med, multimodal clinical 459 monitoring in the emergency department.Scientific Data12, 1094 (2025).460

    2025

  22. [30]

    Wang, W., Mohseni, P ., Kilgore, K. L. & Najafizadeh, L. Pulsedb: A large, cleaned dataset 461 based on mimic-iii and vitaldb for benchmarking cuff-less blood pressure estimation methods. 462 Frontiers in Digital Health4, 1090854 (2023). 463

    2023

  23. [31]

    Chen, X.et al.Racial/ethnic differences in sleep disturbances: the multi-ethnic study of 464 atherosclerosis (mesa).Sleep38, 877–888 (2015). 465

    2015

  24. [32]

    URL https://doi.org/10.466 60508/qjbv-hg78

    Sun, H.et al.The human sleep project (version 2.0) (2023). URL https://doi.org/10.466 60508/qjbv-hg78. 467

    2023

  25. [33]

    Redline, S.et al.The familial aggregation of obstructive sleep apnea.American journal of 468 respiratory and critical care medicine151, 682–687 (1995).469 17

    1995

  26. [34]

    & Schmidt, P

    Reiss, A., Indlekofer, I. & Schmidt, P . Ppg-dalia. UCI Machine Learning Repository (2019). 470 URLhttps://doi.org/10.24432/C53890. 471

    2019 doi

  27. [35]

    & Shabany, M

    Kachuee, M., Kiani, M., Mohammadzade, H. & Shabany, M. Cuff-less blood pressure 472 estimation dataset. UCI Machine Learning Repository (2015). URL https://doi.org/10.473 24432/C5B602. 474

    2015

  28. [36]

    Nemcova, A.et al.Brno university of technology smartphone ppg database (but ppg): 475 Annotated dataset for ppg quality assessment and heart rate estimation.BioMed Research 476 International2021, 3453007 (2021). 477

    2021

  29. [37]

    & Lee, J

    Lee, H., Chung, H. & Lee, J. Motion artifact cancellation in wearable photoplethysmography 478 using gyroscope.IEEE Sensors Journal19, 1166–1175 (2018).479

    2018

  30. [38]

    & Van Laerhoven, K

    Schmidt, P ., Reiss, A., Duerichen, R., Marberger, C. & Van Laerhoven, K. Introducing wesad, 480 a multimodal dataset for wearable stress and affect detection. InProceedings of the 20th 481 ACM international conference on multimodal interaction, 400–408 (2018).482

    2018

  31. [39]

    & Ashley, E

    Torres-Soto, J. & Ashley, E. A. Multi-task deep learning for cardiac rhythm detection in 483 wearable devices.NPJ digital medicine3, 116 (2020). 484

    2020

  32. [40]

    E.et al.Mimic-iii, a freely accessible critical care database.Scientific data3, 485 1–9 (2016)

    Johnson, A. E.et al.Mimic-iii, a freely accessible critical care database.Scientific data3, 485 1–9 (2016). 486

    2016

  33. [41]

    Lee, H.-C.et al.Vitaldb, a high-fidelity multi-parameter vital signs database in surgical 487 patients.Scientific Data9, 279 (2022). 488

    2022

  34. [42]

    On the analysis of fingertip photoplethysmogram signals.Current cardiology 489 reviews8, 14–25 (2012)

    Elgendi, M. On the analysis of fingertip photoplethysmogram signals.Current cardiology 489 reviews8, 14–25 (2012). 490

    2012

  35. [43]

    & Zhang, Y

    Zhao, Z. & Zhang, Y . Sqi quality evaluation mechanism of single-lead ecg signal based on 491 simple heuristic fusion and fuzzy comprehensive evaluation.Frontiers in physiology9, 727 492 (2018). 493

    2018

  36. [44]

    InProceedings of the 26th ACM SIGKDD International Conference on 495 Knowledge Discovery & Data Mining, 1614–1624 (2020).496

    Hong, S.et al.Holmes: Health online model ensemble serving for deep learning models in494 intensive care units. InProceedings of the 26th ACM SIGKDD International Conference on 495 Knowledge Discovery & Data Mining, 1614–1624 (2020).496

    2020

  37. [45]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In 497 Proceedings of the IEEE conference on computer vision and pattern recognition, 770–778 498 (2016). 499

    2016

  38. [46]

    & Sun, G

    Hu, J., Shen, L. & Sun, G. Squeeze-and-excitation networks. InProceedings of the IEEE 500 conference on computer vision and pattern recognition, 7132–7141 (2018).501

    2018

  39. [47]

    Loshchilov, I., Hutter, F .et al.Fixing weight decay regularization in adam.arXiv preprint 502 arXiv:1711.051015, 5 (2017). 503

    2017 arXiv

  40. [48]

    Ndumele, C. E.et al.A synopsis of the evidence for the science and clinical management of 504 cardiovascular-kidney-metabolic (ckm) syndrome: a scientific statement from the american 505 heart association.Circulation148, 1636–1664 (2023). 506

    2023

  41. [49]

    Ndumele, C. E.et al.Cardiovascular-kidney-metabolic health: a presidential advisory from 507 the american heart association.Circulation148, 1606–1635 (2023).508 18 A Diagnostic performance of AnyPPG for ICD-10 diagnoses509 The diagnostic performance of AnyPPG for ICD-10 codes ...

    2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.