Pith. sign in

REVIEW 5 major objections 6 minor 50 references

Photoplethysmography, guided by physiology-based self-supervision, can screen for aortic stenosis and regurgitation with moderate accuracy despite scarce clinical labels.

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

2026-08-03 04:39 UTC pith:YCLX5LAP

load-bearing objection A serious, well-resourced PPG screening paper whose central claim is unproven as written: PiLA is never compared to its own architecture trained from scratch, and the headline AUROC differs between the abstract and the full text. the 5 major comments →

arxiv 2602.04266 v2 pith:YCLX5LAP submitted 2026-02-04 eess.SP cs.LG

Aortic Valve Disease Screening from PPG via Physiology-Guided Self-Supervised Learning

Jiaze Wang , Qinghao Zhao , Zizheng Chen , Zhejun Sun , Deyun Zhang , Yuxi Zhou , Shenda Hong This is my paper
classification eess.SP cs.LG
keywords Aortic valve diseasePhotoplethysmography (PPG)Self-supervised learningAortic stenosisAortic regurgitationPseudo-label pretext taskUK BiobankWearable screening
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.

The paper tries to show that aortic valve disease can be screened from a single averaged peripheral pulse wave (PPG) using a self-supervised model that never sees valve labels during pre-training. Instead, it builds pseudo-labels from textbook waveform signs — the slow upstroke of aortic stenosis, the water-hammer collapse of aortic regurgitation, and the anacrotic notch — and pre-trains on 170,702 unlabeled UK Biobank recordings. After fine-tuning on a small labeled cohort, the model reaches AUROC 0.7645 for AS and 0.7756 for AR (the abstract reports 0.8025 and 0.7669), beating supervised and generic self-supervised baselines. It also claims the model score is an independent predictor of future valve events after adjustment for standard risk factors. The payoff would be a low-cost, scalable screening tool built into the wearables people already wear.

Core claim

PiLA, the proposed physiology-guided self-supervised learning framework, claims that a proxy task defined by four PPG morphological phenotypes (anacrotic, tardus, water-hammer, normal) compels the encoder to learn stable hemodynamic features from unlabeled data. After fine-tuning on 5,460 labeled subjects (245 AS, 213 AR) with a dual-branch gated fusion architecture, it achieves AUROC 0.7645 for AS and 0.7756 for AR, outperforming supervised ResNet1D, TimesNet, Attn-LRCN, and generic SSL (SimCLR, reconstruction, clustering). The paper further claims the output is an independent risk marker: in fully adjusted Cox models, per-SD hazard ratios are 1.33 (AS) and 1.38 (AR), and in PSM-matched coh

What carries the argument

The load-bearing object is the Arterial Pulse Pattern Recognition (APPR) proxy task: pseudo-labels are generated by thresholding the ratio of upstroke time to systolic time (>60% = tardus, <30% = water-hammer) plus algorithmic detection of an anacrotic notch. Pre-training runs on 170,702 unlabeled samples using Multi-Stream ResNet (M-ResNet), which takes PPG plus first and second derivatives (VPG, APG) as parallel input channels to capture inflection points. Fine-tuning uses a dual-branch architecture where a frozen transfer branch carries the pre-trained representation and a scratch-trained supervised branch is filtered by a gating network before fusion.

Load-bearing premise

The premise is that the shape of a single averaged peripheral pulse wave, as captured by the PulseTrace PCA2 device, faithfully reveals the central hemodynamic changes caused by aortic stenosis and regurgitation, despite age-related arterial stiffening and other physiological confounders.

What would settle it

Train PiLA with pseudo-labels randomly shuffled (or based on heart rate alone) and fine-tune on the same labeled set; if AUROC remains at 0.7645/0.7756 or higher, then the physiological guidance is not responsible for the gain.

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

If this is right

  • If the central claim holds, a single averaged fingertip PPG waveform is enough to flag valvular disease risk years before clinical diagnosis, enabling population-scale screening.
  • The same physiology-guided pretext-task recipe could be reused for other valvular or hemodynamic conditions where labeled data are scarce.
  • Practical screening strategies could target the top 5–10% risk scores from the model to concentrate echocardiography resources (enrichment factor up to 4.68x for AS).
  • The model's independence from age, blood pressure, diabetes, and CVD history suggests it adds an orthogonal signal to existing risk scores.

Where Pith is reading between the lines

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

  • Because the pseudo-label rules are derived from the same waveform features used by the UK Biobank device, the pre-training could inadvertently learn device-specific artifacts; external validation on another device or population is needed to confirm generalization.
  • The abstract and results report different AUROC values (0.8025/0.7669 vs 0.7645/0.7756), which may reflect different evaluation settings; readers should check which number applies to prospective use.
  • If the physiology-guided labels are truly the driver, then shuffling or corrupting the pseudo-labels during pre-training should substantially degrade fine-tuned performance — a testable experimental control.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes PG-SSL/PiLA, a self-supervised framework for aortic stenosis (AS) and aortic regurgitation (AR) screening from single-beat averaged PPG waveforms (100 samples, UK Biobank PulseTrace PCA2). The method uses ~170,000 unlabeled PPG recordings to pre-train a multi-stream ResNet (M-ResNet) on a pseudo-label classification task built from hand-defined morphological rules (pulsus tardus, water-hammer pulse, anacrotic notch), followed by fine-tuning on a small labeled cohort (245 AS, 213 AR) with a dual-branch gated-fusion architecture. The paper reports AUROCs of 0.7645 (AS) and 0.7756 (AR) in Table 2, claims significant improvement over supervised baselines, and presents subgroup, calibration, temporal, and survival analyses supporting independent prognostic value.

Significance. If the central claim holds, this is a meaningful contribution: it suggests that large unlabeled PPG collections can be exploited through physiology-derived pretext tasks to build AVD screening tools where echocardiography-based labels are scarce. The paper has clear strengths: it uses a very large real-world cohort, publicly releases code, formulates clinically motivated pseudo-labels explicitly, and includes several secondary analyses (calibration, enrichment, temporal sensitivity, propensity-matched survival). The temporal dose-response and independent prognostic associations are valuable exploratory evidence. However, the primary claim of a pretraining benefit is not yet established because the main comparator differs from PiLA in architecture as well as pretraining, and because several reported numbers are internally inconsistent. The significance of the paper will depend on resolving these issues.

major comments (5)
  1. [Abstract vs. Results/Table 2] The abstract reports AUROCs of 0.8025 (AS) and 0.7669 (AR), while the Results section and Table 2 report 0.7645 (AS) and 0.7756 (AR). Figure 1 reports 0.765/0.776. These are inconsistent and the discrepancy is not acknowledged. The headline result must be a single, internally consistent set of numbers, with exact test-set sizes and, ideally, confidence intervals.
  2. [Table 2 / Fine-tuning comparison] The central claim—that PG-SSL pretraining significantly improves over supervised learning—is not supported by the reported comparison. PiLA differs from the ResNet1D baseline not only by pretraining but also by using multi-stream derivative inputs (PPG/VPG/APG), a dual-branch M-ResNet architecture, a gating fusion head, and a different fine-tuning objective. Without a from-scratch run of the exact PiLA architecture (random initialization, same training schedule, same fusion head), the observed AUC gaps (AS: 0.7645 vs. 0.6985; AR: 0.7756 vs. 0.7013) cannot be attributed to the pretext task. Please add this architecture-matched baseline.
  3. [Statistical significance / uncertainty, test-set size] The primary endpoint AUROCs are reported without confidence intervals or significance tests. Given a 20% test split, the positive sets are approximately 49 AS and 43 AR, so the differences among methods in Table 2 may be within sampling variability. Provide bootstrap or DeLong CIs, and significance tests for the key comparison. The phrase 'significant performance improvement' in the abstract and Results is not currently justified.
  4. [Cohort definitions / discrepant sample sizes] The Methods state the labeled dataset contains 5,460 subjects (245 AS, 213 AR), but the subgroup analysis refers to '6,377 subjects included in the study,' and the PSM cohorts are reported as n=6,786 (AS) and n=3,527 (AR), which exceed the total labeled cohort. The survival analysis in Table 4 is said to use the 'total cohort' but does not define its denominator. These inconsistencies make it difficult to interpret the longitudinal results. Please clarify the cohort construction, event definitions, and which participants enter each analysis.
  5. [Methods: pseudo-label generation / specificity of the pretext task] The pretext pseudo-labels are generated from the same textbook morphological signatures that define AS and AR (UT/ST thresholds, notch detection). While this is not label leakage, it means the claimed advantage of PG-SSL over generic SSL may simply reflect that the pretext task already encodes target-relevant morphology. To support the claim that the physiology-guided rules themselves are responsible for the gains, include an ablation with randomized pseudo-labels or with a pretext task based on the same UT/ST features but shuffled labels; this would also address concerns about threshold arbitrariness.
minor comments (6)
  1. [Title] The title uses 'Physiology-Informed' while the abstract and body use 'Physiology-Guided' (PG-SSL). Please standardize.
  2. [Table 3] Table 3 contains a typo: '1–3 Y ears' should be 'Years'. Also ensure the confidence intervals in Table 3 are computed in the same way as requested for Table 2.
  3. [Table 4] The p-values are printed as '<10−47' etc., which appears to be a formatting artifact; use standard scientific notation.
  4. [Methods / input representation] Methods define x_i in R^L as a 'pre-processed PPG time-series segment,' but the Data section states the UK Biobank device outputs an averaged 100-point waveform. Clarify that L=100 and that no further preprocessing is applied, and whether derivative channels are computed from the same 100-point waveform.
  5. [Hyperparameters / reproducibility] The paper gives no hyperparameter details (epochs, batch size, learning rate, optimizer, number of pretrained samples used, fine-tuning schedule) in the main text or appendix. Since the code is claimed to be public, please include this information or a pointer to a configuration file.
  6. [References] Reference 30 is incomplete (no journal/venue or arXiv identifier). Some references to guidelines use 'ESC/EACTS Task Force (2025)' but no author list; please complete.

Circularity Check

1 steps flagged

Central SSL comparison is not circular; one secondary Grad-CAM 'physiological fidelity' claim is self-confirming because it evaluates attention to the same waveform features used to define the pretext pseudo-labels.

specific steps
  1. self definitional [Results, 'Physiological Fidelity and Robustness Analysis' (Grad-CAM, Figure 2); Methods, 'PPG Morphological Phenotyping for AVD' / 'Pre-training via Pulse Pattern Recognition Task']
    "In the Aortic Stenosis (AS) group, the model exhibited a highly focused attention pattern, primarily targeting the systolic upstroke and showing secondary attention at the dicrotic notch on the descent. The highlighting of the upstroke corresponds to the classic Pulsus Parvus feature of AS... [However, the pretext pseudo-labels were defined as] 'Anacrotic Pulse: Identified by the presence of an algorithmically detected notch (shoulder) on the systolic upstroke...' and 'Pulsus Tardus ... ratio of UT to the total systolic time (ST) exceeds 60%'."

    The model's attention to the systolic upstroke and dicrotic notch is reported as independent confirmation that PiLA learned AS/AR hemodynamics. But those are exactly the cues (notch presence; UT/ST ratio) used to construct the four pseudo-labels in the pre-training task. The Grad-CAM result therefore verifies the pretext rule-set, not a newly discovered physiological correspondence. This circularity is limited to the interpretation of the attention maps; the main fine-tuning AUC results are against ICD-10-based AS/AR labels and do not reduce to the pseudo-label definitions.

full rationale

The paper's central empirical claim—that pretraining on physiology-derived pseudo-labels improves downstream AS/AR screening AUC relative to supervised and generic-SSL baselines—is not circular by construction. Pseudo-labels are generated from fixed literature-derived thresholds (UT/ST >60%, <30%, notch detection), not from the downstream AS/AR ICD labels or from the test set; no fitted parameter is renamed as a prediction. There are no load-bearing self-citations: the references defining the thresholds (Yoshioka et al., Boiteau et al., Meghraoui et al.) are external, and no central claim rests on the authors' prior work. The main caveat is architectural: PiLA differs from the ResNet1D baseline in input streams, dual-branch gating, and fine-tuning objective, so the gain cannot be uniquely attributed to the pretext task—but this is a confound in the experimental comparison, not circularity. The abstract/Results AUROC discrepancy and absence of confidence intervals are also correctness/consistency concerns, not circularity. The only partially self-confirming step is the Grad-CAM 'physiological fidelity' narrative, which re-derives the pretext task's own definitions; hence a low score of 2 rather than 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are proposed. The model's free parameters are the hand-chosen pseudo-label thresholds and unspecified measurement algorithms; the axioms are the physiological transfer assumptions from central to peripheral pulse waves and the label accuracy assumptions.

free parameters (5)
  • Tardus pseudo-label threshold (UT/ST > 60%) = 60%
    Chosen from clinical literature but adjusted by authors ('to enhance specificity'); affects which unlabeled samples are labeled Tardus and thus the entire pretext task.
  • Water-hammer pseudo-label threshold (UT/ST < 30%) = 30%
    Chosen from literature (Boiteau et al. suggested <34%) and made stricter; defines WHP pseudo-class.
  • Positive-case diagnostic window (diagnosis within 1 year after PPG) = 1 year
    Design choice that defines the labeled cohort; could introduce misclassification if some 'positive' patients did not have disease at PPG time.
  • UT/ST measurement algorithm
    The method for detecting systolic onset, peak, and dicrotic notch on the 100-point averaged waveform is not specified; different implementations change pseudo-labels.
  • Anacrotic notch detection threshold
    Rule for detecting a shoulder on the upstroke is referenced to Fleming but no algorithmic threshold is provided.
axioms (5)
  • domain assumption Peripheral PPG morphology reliably reflects central aortic pressure waveform dynamics (ref 12).
    Load-bearing: if false, AS/AR hemodynamic signatures are not present in the PPG waveform.
  • domain assumption Pulsus tardus/parvus and water-hammer pulse phenotypes are specific enough to AS/AR to serve as pseudo-labels.
    The pretext task assumes these waveform phenotypes are disease-relevant, despite overlap with age and arterial stiffness.
  • domain assumption UK Biobank unlabeled participants without recorded AVD labels are an appropriate pretraining population.
    Unlabeled set may contain undiagnosed AVD, but SSL is robust to pseudo-label noise; still an assumption.
  • domain assumption ICD-10 codes, self-report, and the 1-year diagnostic window identify true AS/AR status.
    No echocardiography adjudication is available; misclassification would attenuate or bias performance estimates.
  • standard math Standard deep learning fine-tuning assumptions (i.i.d. splits, no leakage between pretrain and fine-tune).
    The paper splits by subject but does not state whether pretraining subjects overlap with labeled subjects; UKB participants are unique, so likely satisfied.

pith-pipeline@v1.3.0-alltime-deepseek · 19312 in / 12942 out tokens · 132131 ms · 2026-08-03T04:39:51.295261+00:00 · methodology

0 comments
read the original abstract

Aortic valve disease (AVD) represents a major public health burden, while its diagnosis relies on echocardiography, which is limited by cost and specialist expertise, restricting scalable screening and risk stratification. Existing portable sensing modalities are constrained by indirect representations or acquisition dependencies. In this context, photoplethysmography (PPG), a widely available optical signal capturing peripheral hemodynamic dynamics, provides a scalable physiological measurement. However, the scarcity of clinically labeled PPG data severely constrains the development of effective data-driven models. To address this limitation, we propose Physiology-Guided Self-Supervised Learning (PG-SSL), leveraging approximately 170,000 unlabeled UK Biobank PPG recordings. PG-SSL constructs physiologically derived pseudo-labels based on clinically motivated waveform phenotypes associated with aortic stenosis (AS) and aortic regurgitation (AR), enabling large-scale pretraining without AVD-specific labels. Following fine-tuning on a small labeled cohort, the model achieved AUROCs of 0.8025 for AS and 0.7669 for AR. Further analyses demonstrated robustness under clinical confounding and covariate-balanced evaluation, as well as significant longitudinal associations with incident AVD events. This study demonstrates the feasibility of PG-SSL for leveraging large-scale unlabeled physiological signals under clinically labeled data-scarce conditions. The proposed approach provides a useful strategy for improving low-cost PPG-based screening and risk enrichment for clinically recognized AVD.

Figures

Figures reproduced from arXiv: 2602.04266 by Deyun Zhang, Jiaze Wang, Qinghao Zhao, Shenda Hong, Yuxi Zhou, Zhejun Sun, Zizheng Chen.

Figure 1
Figure 1. Figure 1: Performance evaluation of the model on the independent test set. (A, B) ROC curves [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Grad-CAM visualization of model attention across different groups. (Left) Healthy [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Temporal sensitivity analysis of the model. (A, B) ROC curves stratified by time-to [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Kaplan-Meier survival analysis in the PSM-matched cohort. ( [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Forest plots of subgroup analysis for (a) Aortic Stenosis and (b) Aortic Regurgitation. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The overall architecture of the PiLA framework. The workflow illustrates the process [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Detailed architecture of the deep learning model. [PITH_FULL_IMAGE:figures/full_fig_p021_7.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

50 extracted references · 16 canonical work pages

  1. [1]

    Coffey, S., Roberts-Thomson, R., Brown, A., Carapetis, J., Chen, M., Enriquez-Sarano, M., Z¨uhlke, L., and Prendergast, B.D. (2021). Global epidemiology of valvular heart disease. Nature Reviews Cardiology18, 853–864. doi:10.1038/s41569-021-00570-z

    work page doi:10.1038/s41569-021-00570-z 2021

  2. [2]

    Aluru, J.S., Barsouk, A., Saginala, K., Rawla, P ., and Barsouk, A. (2022). Valvular heart disease epidemiology. Medical Sciences10, 32. doi:10.3390/medsci10020032

    work page doi:10.3390/medsci10020032 2022

  3. [3]

    Thoenes, M., Bramlage, P ., Zamorano, P ., Messika-Zeitoun, D., Wendt, D., Kasel, M., Ku- rucova, J., and Steeds, R.P . (2018). Patient screening for early detection of aortic stenosis (as)–review of current practice and future perspectives. Journal of Thoracic Disease10, 5584–5594. doi:10.21037/jtd.2018.09.02

    work page doi:10.21037/jtd.2018.09.02 2018

  4. [4]

    2025 esc/eacts guidelines for the management of valvular heart disease

    ESC/EACTS Task Force (2025). 2025 esc/eacts guidelines for the management of valvular heart disease. European Heart Journal. doi:10.1093/eurheartj/ehaf194

    work page doi:10.1093/eurheartj/ehaf194 2025

  5. [5]

    Khade, P .J., Mane, P ., Mahore, S., and Bhole, K. (2021). Machine learning approach for prediction of aortic and mitral regurgitation based on phonocardiogram signal. In 2021 12th International Conference on Computing Communication and Networking Technologies (ICCCNT) (IEEE, 2021). . doi:10.1109/ICCCNT51525.2021.9579971

    arXiv 2021

  6. [6]

    Erin, E., and Semiz, B. (2024). Performance comparison of gyrocardiogram and seismo- cardiogram signals in valvular heart disease assessment. In BIOSTEC 2024: Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Tech- nologies (SCITEPRESS, 2024). . doi:10.5220/0012441700003657

    work page doi:10.5220/0012441700003657 2024

  7. [7]

    Saraf, K., Baek, C.I., Wasko, M.H., Zhang, X., Zheng, Y ., Borgstrom, P .H., Mahajan, A., and Kaiser, W.J. (2019). Fully-automated diagnosis of aortic stenosis using phonocardiogram- based features. In 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) (IEEE, 2019). . doi:10.1109/EMBC.2019.8857506

    arXiv 2019

  8. [8]

    Cohen-Shelly, M., Attia, Z.I., Friedman, P .A., Ito, S., Essayagh, B.A., Ko, W.Y ., Murphree, D.H., Michelena, H.I., Enriquez-Sarano, M., and Carter, R.E. (2021). Electrocardiogram screening for aortic valve stenosis using artificial intelligence. European Heart Journal42, 2885–2896. doi:10.1093/eurheartj/ehab153

    doi:10.1093/eurheartj/ehab153 2021

  9. [9]

    Sawano, S., Kodera, S., Katsushika, S., Nakamoto, M., Ninomiya, K., Shinohara, H., Hi- gashikuni, Y ., Nakanishi, K., Nakao, T., and Seki, T. (2022). Deep learning model to detect significant aortic regurgitation using electrocardiography. Journal of Cardiology79, 334–

    2022

  10. [10]

    Pereira, T., Tran, N., Gadhoumi, K., Pelter, M.M., Do, D.H., Lee, R.J., Colorado, R., Meisel, K., and Hu, X. (2020). Photoplethysmography based atrial fibrillation detection: a review. npj Digital Medicine3, 3. doi:10.1038/s41746-019-0207-9

    work page doi:10.1038/s41746-019-0207-9 2020

  11. [11]

    Allen, J. (2007). Photoplethysmography and its application in clinical physiological measure- ment. Physiological Measurement28, R1–R39. doi:10.1088/0967-3334/28/3/R01

    work page doi:10.1088/0967-3334/28/3/r01 2007

  12. [12]

    Millasseau, S.C., Ritter, J.M., Takazawa, K., and Chowienczyk, P .J. (2006). Contour analysis of the photoplethysmographic pulse measured at the finger. Journal of Hypertension24, 1449–1456. doi:10.1097/01.hjh.0000239277.05068.87

    arXiv 2006

  13. [13]

    Elgendi, M., Fletcher, R., Liang, Y ., Howard, N., Lovell, N.H., Abbott, D., Lim, K., and Ward, R. (2019). The use of photoplethysmography for assessing hypertension. npj Digital Medicine2, 60. doi:10.1038/s41746-019-0136-7

    doi:10.1038/s41746-019-0136-7 2019

  14. [14]

    Karimpour, P ., May, J.M., and Kyriacou, P .A. (2023). Photoplethysmography for the assess- ment of arterial stiffness. Sensors23, 9882. doi:10.3390/s23249882

    doi:10.3390/s23249882 2023

  15. [15]

    Y ang, R., Kang, L., Zhou, A., Cui, H., and Ma, H. (2024). Aortic stenosis detection by improved inception convolution network-enabled pulse wave. In 2024 IEEE 30th Inter- national Conference on Parallel and Distributed Systems (ICPADS) (IEEE, 2024). . doi: 10.1109/ICPADS63350.2024.00024

    arXiv 2024

  16. [16]

    Kang, L., Y ang, R., Liu, Y ., Zhou, A., Ma, H., and Cui, H. (2025). PASS: A novel PPG-based aortic stenosis screening system. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies9, 1–25. doi:10.1145/3749548

    work page doi:10.1145/3749548 2025

  17. [17]

    Krishnan, R., Rajpurkar, P ., and Topol, E.J. (2022). Self-supervised learning in medicine and healthcare. Nature Biomedical Engineering6, 1346–1352. doi:10.1038/ s41551-022-00914-1

    2022

  18. [18]

    Zhang, K., Wen, Q., Zhang, C., Cai, R., Jin, M., Liu, Y ., Zhang, J.Y ., Liang, Y ., Pang, G., and Song, D. (2024). Self-supervised learning for time series analysis: Taxonomy, progress, and prospects. IEEE Transactions on Pattern Analysis and Machine Intelligence46, 6775–6794. doi:10.1109/TPAMI.2024.3387317

    arXiv 2024

  19. [19]

    McGee, S. (2021). Evidence-Based Physical Diagnosis. 5 ed.. Elsevier

    2021

  20. [20]

    Wu, H., Hu, T., Liu, Y ., Zhou, H., Wang, J., and Long, M. (2022). Timesnet: Temporal 2d-variation modeling for general time series analysis. arXiv preprint arXiv:2210.02186

    Pith/arXiv arXiv 2022

  21. [21]

    Choi, J., Hwang, G., Lee, J.S., Ryu, M., and Lee, S.J. (2023). Weighted knowledge distilla- tion of attention-lrcn for recognizing affective states from ppg signals. Expert Systems with Applications233, 120883. doi:10.1016/j.eswa.2023.120883

    arXiv 2023

  22. [22]

    Ding, C., and Wu, C. (2024). Self-supervised learning for biomedical signal processing: A systematic review on ECG and PPG signals. medRxiv. . doi:10.1101/2024.09.30. 24314588

    work page doi:10.1101/2024.09.30 2024

  23. [23]

    Weng, W., Gu, Y ., Guo, S., Ma, Y ., Y ang, Z., Liu, Y ., and Chen, Y . (2025). Self-supervised learning for electroencephalogram: A systematic survey. ACM Computing Surveys57, 1–

    2025

  24. [24]

    Liu, Z., Alavi, A., Li, M., and Zhang, X. (2023). Self-supervised contrastive learning for medical time series: A systematic review. Sensors23, 4221. doi:10.3390/s23094221. 24

    work page doi:10.3390/s23094221 2023

  25. [25]

    Banville, H., Chehab, O., Hyv ¨arinen, A., Engemann, D.A., and Gramfort, A. (2021). Uncov- ering the structure of clinical EEG signals with self-supervised learning. Journal of Neural Engineering18, 046020. doi:10.1088/1741-2552/abca18

    doi:10.1088/1741-2552/abca18 2021

  26. [26]

    Mehari, T., and Strodthoff, N. (2022). Self-supervised representation learning from 12-lead ECG data. Computers in Biology and Medicine141, 105114. doi:10.1016/j.compbiomed. 2021.105114

    arXiv 2022

  27. [27]

    von Rueden, L., Mayer, S., Beckh, K., Georgiev, B., Giesselbach, S., Heese, R., Kirsch, B., Pfrommer, J., Pick, A., Ramamurthy, R., Walczak, M., Garcke, J., Bauckhage, C., and Schuecker, J. (2023). Informed machine learning: A taxonomy and survey of integrating prior knowledge into learning systems. IEEE Transactions on Knowledge and Data Engi- neering35,...

    arXiv 2023

  28. [28]

    Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P ., Wang, S., and Y ang, L. (2021). Physics-informed machine learning. Nature Reviews Physics3, 422–440. doi:10.1038/ s42254-021-00314-5

    2021

  29. [29]

    Liu, R., Shen, J., Gu, Y ., Chen, Y ., Zhang, J., Wu, Q., Xu, C., and Fan, F . (2025). Physcl: Knowledge-aware contrastive learning of physiological signal models for cuff-less blood pressure estimation. IEEE Journal of Biomedical and Health Informatics29, 4728–4740. doi:10.1109/JBHI.2025.3554495

    arXiv 2025

  30. [30]

    Maghsoodi, N., Nassar, S., Wilson, P .F .R., To, M.N.N., Mannina, S., Addas, S., Sibley, S., Maslove, D., Abolmaesumi, P ., and Mousavi, P . (2025). Domain knowledge is power: Lever- aging physiological priors for self-supervised representation learning in electrocardiography

    2025

  31. [31]

    Lee, B.T., Kong, S.T., Song, Y ., and Lee, Y . (2021). Self-supervised learning with electrocar- diogram delineation for arrhythmia detection. In 2021 43rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) (IEEE, 2021). . doi: 10.1109/EMBC46164.2021.9630364

    arXiv 2021

  32. [32]

    Meghraoui, M.H., Benaired, N., Benselama, Z.A., Yssaad, B., and Benselama, S.I. (2024). Classifying abnormal arterial pulse patterns in cardiovascular diseases: A photoplethys- mography and machine learning approach. Traitement du Signal41, 543–562. doi: 10.18280/ts.410210

    work page doi:10.18280/ts.410210 2024

  33. [33]

    Abbas, A.E., and Pibarot, P . (2019). Hemodynamic characterization of aortic stenosis states. Catheterization and Cardiovascular Interventions93, 1002–1023. doi:10.1002/ ccd.28146

    2019

  34. [34]

    Hungerford, S.L., Adji, A.I., and Kapur, N.K. (2023). Interpretation of the central aortic pres- sure waveform in elderly patients with aortic stenosis. American Journal of Physiology - Heart and Circulatory Physiology324, H697–H712. doi:10.1152/ajpheart.00050.2023

    arXiv 2023

  35. [35]

    Piccioli, F ., Li, Y ., Valiani, A., Caleffi, V., Chowienczyk, P ., and Alastruey, J. (2023). Cardiac contractility is a key factor in determining pulse pressure and its peripheral amplification. Frontiers in Cardiovascular Medicine10, 1197842. doi:10.3389/fcvm.2023.1197842

    arXiv 2023

  36. [36]

    Yu, S., and McEniery, C.M. (2020). Central versus peripheral artery stiffening and car- diovascular risk. Arteriosclerosis, Thrombosis, and Vascular Biology40, 1028–1033. doi: 10.1161/ATVBAHA.120.313128. 25

    work page doi:10.1161/atvbaha.120.313128 2020

  37. [37]

    Elgendi, M. (2012). On the analysis of fingertip photoplethysmogram signals. Current Car- diology Reviews8, 14–25. doi:10.2174/157340312801215782

    doi:10.2174/157340312801215782 2012

  38. [38]

    Saeed, S., Dweck, M.R., and Chambers, J. (2020). Sex differences in aortic stenosis: from pathophysiology to treatment. Expert Review of Cardiovascular Therapy18, 65–76. doi: 10.1080/14779072.2020.1732209

    arXiv 2020

  39. [39]

    Pedroso, A.F ., and Khera, R. (2025). Leveraging ai-enhanced digital health with consumer devices for scalable cardiovascular screening, prediction, and monitoring. npj Cardiovascu- lar Health2, 34. doi:10.1038/s44325-025-00071-9

    doi:10.1038/s44325-025-00071-9 2025

  40. [40]

    Wong, K.C., Nguyen, T.N., and Chow, C.K. (2024). Global implementation and evaluation of atrial fibrillation screening in the past two decades–a narrative review. npj Cardiovascular Health1, 17. doi:10.1038/s44325-024-00014-w

    work page doi:10.1038/s44325-024-00014-w 2024

  41. [41]

    Chen, W., Khurshid, S., Singer, D.E., Atlas, S.J., Ashburner, J.M., Ellinor, P .T., Mc- Manus, D.D., Lubitz, S.A., and Chhatwal, J. (2022). Cost-effectiveness of screening for atrial fibrillation using wearable devices. JAMA Health Forum3, e222419. doi:10.1001/ jamahealthforum.2022.2419

    arXiv 2022

  42. [42]

    Azizi, S., Culp, L., Freyberg, J., Mustafa, B., Baur, S., Kornblith, S., Chen, T., Toma- sev, N., Mitrovi ´c, J., Strachan, P ., Mahdavi, S.S., Wulczyn, E., Babenko, B., Walker, M., Loh, A., Chen, P .H.C., Liu, Y ., Bavishi, P ., McKinney, S.M., Winkens, J., Guha Roy, A., Beaver, Z., Ryan, F ., Krogue, J.D., Etemadi, M., Telang, U., Liu, Y ., Peng, L., Co...

    2023

  43. [43]

    Ghorbani, R., Reinders, M.J.T., and Tax, D.M.J. (2023). Self-supervised ppg representa- tion learning shows high inter-subject variability. In Proceedings of the 8th International Conference on Machine Learning Technologies (ICMLT) (2023)

    2023

  44. [44]

    Majeed, A., and Hwang, S.O. (2024). A data-centric ai paradigm for socio-industrial and global challenges. Electronics13, 2156. doi:10.3390/electronics13112156

    work page doi:10.3390/electronics13112156 2024

  45. [45]

    Moxham, I.M. (2003). Understanding arterial pressure waveforms. Southern African Journal of Anaesthesia and Analgesia9, 40–42. doi:10.1080/22201173.2003.10872991

    arXiv 2003

  46. [46]

    Suvarna, J. (2008). Watson’s water hammer pulse. Journal of Postgraduate Medicine54, 163

    2008

  47. [47]

    Fleming, P .R. (1957). The mechanism of the pulsus bisferiens. British Heart Journal19, 519–524

    1957

  48. [48]

    Y oshioka, N., Fujita, Y ., Y asukawa, T., Sano, I., Kiso, M., Nakayama, M., and Y oshida, Y . (2010). Do radial arterial pressure curves have diagnostic validity for identify severe aortic stenosis? Journal of Anesthesia24, 7–10. doi:10.1007/s00540-009-0837-1

    work page doi:10.1007/s00540-009-0837-1 2010

  49. [49]

    Boiteau, G.M., Libanoff, A.J., and Allenstein, B.J. (1964). Upstroke time ratio in valvu- lar aortic insufficiency. The American Journal of Cardiology14, 162–168. doi:10.1016/ 0002-9149(64)90127-4. 26 APPENDIX Definition of Handcrafted PPG Features Table S5 lists the 22 handcrafted PPG morphological features used in the late fusion experiment. These featu...

    1964

  50. [341]

    doi:10.1016/j.jjcc.2021.08.029. 23

    work page doi:10.1016/j.jjcc.2021.08.029 2021