Pith. sign in

REVIEW 4 major objections 6 minor 12 references

A curriculum-adversarial learning framework on four-wavelength PPG estimates systolic blood pressure with 14.2 mmHg mean absolute error and diastolic with 6.4 mmHg on subjects strictly held out from training.

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 ↗

Combining hypertension classification as a curriculum and subject-identity adversarial training improves multi-wavelength PPG blood pressure regression under subject-level splitting, reporting SBP MAE 14.2 mmHg and DBP MAE 6.4 mmHg.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Subject-level splitting is the right call and the paper gets that right; everything else is a modest, plausibly real but statistically under-supported gain. the 4 major comments →

arxiv 2509.12518 v1 pith:7BGW22WQ submitted 2025-09-15 eess.SP cs.LG

Generalizable Blood Pressure Estimation from Multi-Wavelength PPG Using Curriculum-Adversarial Learning

Zequan Liang , Ruoyu Zhang , Wei Shao , Mahdi Pirayesh Shirazi Nejad , Ehsan Kourkchi , Setareh Rafatirad , Houman Homayoun This is my paper
classification eess.SP cs.LG
keywords blood pressure estimationphotoplethysmographymulti-wavelength PPGcurriculum learningdomain-adversarial trainingsubject-level splittingwearable health monitoringdeep learning
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 reading

Blood pressure can be estimated from fingertip light signals (PPG), but earlier models often leaked subject identity by mixing segments from the same person into train and test, inflating accuracy. This paper enforces a strict subject-level split on a public four-wavelength PPG dataset and proposes a two-part training strategy: first classify hypertension status, then gradually regress SBP/DBP, while a gradient-reversed discriminator tries to identify the subject and is fooled. Under this protocol the model reports SBP MAE 14.2 mmHg and DBP MAE 6.4 mmHg, the lowest among all tested baselines, and ablations show both the curriculum and adversarial components matter. A sympathetic reader would care because cuffless, continuous BP monitoring from wearables requires models that transfer to people never seen in training, and this work is a step toward reporting that performance honestly.

Core claim

The paper's central claim is that subject-invariant multi-wavelength PPG features for blood pressure can be learned by combining curriculum learning with domain-adversarial training. The model extracts normalized PPG, derivatives, and envelopes from four wavelengths, fuses them by attention, and feeds three heads: a BP regressor, a hypertension classifier, and a subject-identity discriminator with a gradient-reversal layer. The loss weight λ1 climbs from 0 to 1 across epochs, shifting from coarse hypertension classification to fine SBP/DBP regression, with λ2=1 holding adversarial pressure constant. Under strict subject-level 4:1 splitting of 180 subjects, it reports SBP MAE 14.2 mmHg and DB

What carries the argument

The carrying mechanism is the curriculum-adversarial objective over a shared fused representation. A multi-channel attention layer soft-weights feature vectors from four PPG wavelengths; the fused vector feeds three heads simultaneously. The regression head uses MSE, the curriculum head uses BCE on hypertension status, and the discriminator uses cross-entropy on subject identity, with a gradient-reversal layer that flips the discriminator's gradients during backpropagation. The total loss is L = λ1 L_reg + (1−λ1)L_cls − λ2 L_adv, with λ1 = epoch/total_epochs producing a coarse-to-fine curriculum and λ2 fixed. The adversarial term is what forces the CNNs to drop subject-specific patterns; the

Load-bearing premise

The paper treats the two cuff readings that follow each 30-second PPG recording as ground truth for the whole segment and assumes the 180 recruited subjects are representative enough that a model trained on them will generalize to unseen people.

What would settle it

Run the trained model on a new four-wavelength PPG dataset collected with a different sensor and a wider BP range, with an arterial-line reference; if SBP MAE rises above roughly 15–20 mmHg or errors concentrate at the BP extremes, the generalization claim fails. Independently, log the subject-identity discriminator's accuracy during training: if it stays above chance (e.g., much better than 1/180), the adversarial mechanism is not actually removing subject-specific information.

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

If this is right

  • Strict subject-level evaluation becomes the expected standard: the reported errors are more honest estimates of how the model will fare on people it has never seen than segment-level splits would be.
  • Multi-wavelength fusion is a genuine gain: the four-channel model beats every single wavelength for both SBP and DBP.
  • The adversarial subject-identity loss carries most of the improvement—removing it (with or without the classifier) raises SBP MAE from 14.2 to 16.1–17.1 mmHg—while the curriculum classifier adds a smaller refinement.
  • The reported accuracy is sufficient for hypertension screening and long-term trend monitoring, not for clinical diagnosis, as the paper concludes.

Where Pith is reading between the lines

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

  • The paper never reports the discriminator's subject-identification accuracy; if that accuracy stays far above chance after training, the features are not truly subject-invariant, and the gains could stem from the auxiliary tasks acting as regularizers.
  • Because the public dataset contains only 180 subjects and a BP distribution that is mostly normotensive, the 14.2/6.4 mmHg numbers are an optimistic upper bound; external validation on a larger, more diverse cohort is the obvious next test.
  • The same curriculum-adversarial recipe should transfer to other physiological estimation problems—blood glucose, stress load, oxygen saturation—where subject identity is a known confounder and a coarse clinical label is available.
  • An untested variant would anneal the adversarial weight λ2 over training or make it subject-adaptive; the fixed λ2=1 may over-constrain the encoder late in training and slightly limit regression accuracy.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 6 minor

Summary. The paper proposes a blood-pressure (BP) estimation framework for four-wavelength PPG signals that combines curriculum learning (starting from hypertension classification and gradually shifting to SBP/DBP regression) with domain-adversarial training (confusing subject identity). The authors enforce subject-level splitting on the public Cui et al. dataset (180 subjects) and report SBP MAE 14.2 mmHg and DBP MAE 6.4 mmHg, claiming that the curriculum-adversarial method outperforms existing baselines and that ablations confirm the contribution of both components.

Significance. If the empirical claims withstand statistical scrutiny, the paper offers a useful contribution: it highlights a common leakage problem in prior PPG-based BP studies, correctly applies subject-level splitting, and demonstrates that fusing four wavelengths can improve cross-subject regression performance relative to a strong CNN backbone. The proposed curriculum stage and adversarial subject-ID regularizer are plausible and well motivated. However, the paper is empirical and currently rests on a single subject-level split of a small dataset with no confidence intervals, repeated splits, or paired tests; the reported improvements are small and may be within sampling noise. The authors also provide no code or detailed hyperparameter settings, limiting reproducibility. The strengths are the clear problem framing, the public dataset, and the explicit emphasis on avoiding data leakage.

major comments (4)
  1. [Section III-A and Tables I-III] The central empirical claim rests on a single subject-level 4:1 split of 180 subjects. The reported BHS percentages imply a test set of 33 subjects (e.g., 24.2% = 8/33, 51.5% = 17/33), not the 36 expected from 4:1 splitting; no explanation is given. With n≈33, the differences against the Multi-CNN backbone (SBP MAE 14.2 vs. 16.1, DBP 6.4 vs. 7.2) and the ablation differences (SBP 14.2 vs. 14.3, DBP 6.4 vs. 7.1) are within plausible sampling error for per-subject absolute errors of this magnitude. No confidence intervals, repeated splits, paired tests, or per-subject error distributions are reported. Please provide repeated subject-level splits (e.g., 5 or 10 random 4:1 partitions) with mean±SD or a per-subject error analysis, and reconcile the 33 vs. 36 test-subject discrepancy.
  2. [Abstract, Section III-C, and Section IV] There is an internal contradiction in the headline claim. The Abstract and Section III-C state that the method 'outperforms all baselines' and 'consistently achieves the best performance across all evaluation metrics,' but Section IV states that 'the model does not consistently outperform all baselines across every evaluation metric.' In Table II, This Study ties with CNN1D on DBP BHS-5 (51.5%) and with Multi-CNN on SBP BHS-15 (57.6%). Please align the claims with the actual table entries and with the statistical uncertainty discussed above.
  3. [Section II-A and III-A] The phrase 'generalizable BP estimation' in the title and abstract is stronger than what the experimental design can support. Subject-level splitting removes within-subject leakage, but the dataset contains 180 subjects, one 30-s PPG segment per subject, and two cuff reference readings per subject. No demographic breakdown, signal-quality criteria, or test-set BP distribution is provided, so the reported MAEs may not transfer to other populations or measurement conditions. Please temper the generalizability claim, add external validation if possible, or at least discuss how the BP distribution in Fig. 2 and the unknown demographics limit the scope of the conclusion.
  4. [Section II-E and III-A] The curriculum schedule is under-specified and the two descriptions are inconsistent. Section II-E says the shift begins 'once the classification loss stabilizes,' but Section III-A defines λ1 as the ratio of current epoch to total epochs, with λ1 initialized to 0 and increased linearly. It is also unclear whether λ2 is fixed at 1 throughout or selected by validation. No sensitivity analysis of λ2 or the schedule is reported. Without these details, the specific contribution of the curriculum component cannot be isolated or reproduced. Please specify the exact schedule, the value of λ2, and the model-selection/early-stopping procedure.
minor comments (6)
  1. [Fig. 1] 'Envolope' should be 'Envelope'; the 'train test' annotation under the pipeline is unclear and should be replaced with a proper legend.
  2. [Eq. (4) and Section III-A] Clarify whether λ2 is always 1 or is a tuned hyperparameter; the text says 'initialize the loss weights as λ1=0 and λ2=1' but also calls λ2 a fixed weight.
  3. [Section III-A] 'We perform 5-fold cross-validation within the training set' is mentioned, but no results or usage are described. State how the validation fold is used (e.g., early stopping, hyperparameter selection).
  4. [Section II-A and III-A] The extraction of the 'middle 30 seconds' is not justified. If one segment per subject is used, state this explicitly and explain how the two reference BP readings are averaged or assigned.
  5. [Table II] State explicitly whether all baselines are re-trained and evaluated under the same subject-level splitting and preprocessing pipeline; if some numbers are taken from the original papers, specify this, because protocol differences affect comparability.
  6. [References] Reference [8] is about automated curriculum learning for neural networks; a foundational curriculum-learning reference (e.g., Bengio et al., 2009) may be more appropriate for the proposed coarse-to-fine schedule.

Circularity Check

0 steps flagged

No significant circularity; the only self-citation is motivational and not load-bearing.

full rationale

The paper's central claim is empirical: with subject-level splitting on the public four-wavelength PPG dataset [6], curriculum-adversarial training on a multi-channel CNN improves SBP/DBP regression compared with the Multi-CNN backbone and other baselines. This claim is supported by external benchmarks (A-BiLSTM [6], MLP/CNN1D [12], Multi-CNN [9]) and the reported MAE/BHS metrics, not by a derivation that reduces to its own inputs. The hypertension classifier in Section II-E uses labels thresholded from the same SBP/DBP values used for regression, but this is an explicit auxiliary curriculum task; the classifier is discarded at test time and no hypertension accuracy is reported as a prediction, so no fitted input is renamed as a prediction. The adversarial discriminator uses subject IDs only during training as a regularizer and is likewise removed at test. The sole self-citation is [10] in Section II-D ('Inspired by unsupervised domain adaptation [10], we treat different subjects as distinct domains'); it is motivational, while the actual DANN mechanism is adopted from the external reference [7]. The preprocessing and backbone are from external [9]. The Discussion's caveat that 'the model does not consistently outperform all baselines across every evaluation metric' and the small subject-level test set are statistical/correctness concerns, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The paper adds no new physical entities. It rests on domain assumptions about PPG-BP relationships, label accuracy, and the effectiveness of adversarial subject-invariance. The free parameters are mostly training-schedule and preprocessing choices that are not ablated. All model weights are learned from data, but no code, seeds, or confidence intervals are provided.

free parameters (4)
  • lambda_2 (adversarial loss weight) = 1 (fixed; no sensitivity analysis reported)
    Set at initialization and described as fixed in Section II-E/III-A; the paper gives no ablation over this value, so the reported MAE depends on an unexamined hand choice.
  • lambda_1 curriculum schedule = epoch / total_epochs, starting at 0
    Adaptive schedule in Eq. (4); 'once the classification loss stabilizes' is vague (Section II-E), and no sensitivity analysis is provided.
  • extracted segment length = 30 seconds (middle of each recording)
    Section III-A selects the middle 30 s with no ablation or justification; different segments could change the numbers.
  • hypertension threshold for curriculum labels = SBP>=130 or DBP>=90 mmHg
    Used to create the auxiliary classification targets from the same BP values (Section II-E); a clinical definition, but it determines the curriculum task.
axioms (6)
  • domain assumption Multi-wavelength PPG channels provide complementary BP-related features
    Introduction and Section II-C; the fusion gain in Table I supports but does not prove this generally.
  • domain assumption Subject-level splitting with no overlap yields a fair generalization estimate
    Section II-A and III-A; it reduces leakage but does not guarantee generalization to new populations.
  • domain assumption Adversarial training against subject identity induces subject-invariant features that improve BP regression
    Section II-D, Eq. (1), borrowed from DANN [7]; not directly measured.
  • domain assumption Hypertension classification is an effective curriculum for BP regression
    Section II-E, Eq. (4); ablation partly contradicts this: cls-only worsens SBP MAE.
  • domain assumption The public dataset's BP reference values are accurate labels for the extracted PPG segments
    Section II-A; two cuff measurements per subject are treated as ground truth.
  • domain assumption The preprocessing pipeline from [9] transfers to this dataset
    Section II-B; backbone and preprocessing are adopted from a method developed on a private dataset.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalizable Blood Pressure Estimation from Multi-Wavelength PPG Using Curriculum-Adversarial Learning." pith.science (2026). https://pith.science/paper/7BGW22WQ

@misc{pith2026250912518,
  author       = {Pith},
  title        = {Pith review of: Generalizable Blood Pressure Estimation from Multi-Wavelength PPG Using Curriculum-Adversarial Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7BGW22WQ}},
  note         = {Machine review of arXiv:2509.12518}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Accurate and generalizable blood pressure (BP) estimation is vital for the early detection and management of cardiovascular diseases. In this study, we enforce subject-level data splitting on a public multi-wavelength photoplethysmography (PPG) dataset and propose a generalizable BP estimation framework based on curriculum-adversarial learning. Our approach combines curriculum learning, which transitions from hypertension classification to BP regression, with domain-adversarial training that confuses subject identity to encourage the learning of subject-invariant features. Experiments show that multi-channel fusion consistently outperforms single-channel models. On the four-wavelength PPG dataset, our method achieves strong performance under strict subject-level splitting, with mean absolute errors (MAE) of 14.2mmHg for systolic blood pressure (SBP) and 6.4mmHg for diastolic blood pressure (DBP). Additionally, ablation studies validate the effectiveness of both the curriculum and adversarial components. These results highlight the potential of leveraging complementary information in multi-wavelength PPG and curriculum-adversarial strategies for accurate and robust BP estimation.

Figures

Figures reproduced from arXiv: 2509.12518 by Ehsan Kourkchi, Houman Homayoun, Mahdi Pirayesh Shirazi Nejad, Ruoyu Zhang, Setareh Rafatirad, Wei Shao, Zequan Liang.

Figure 1
Figure 1. Figure 1: Overview of blood pressure estimation from multi-wavelength PPG [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Blood pressure distribution (a) SBP (b) DBP [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Data preprocessing and multi-channel feature extraction [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Multi-channel fusion and curriculum-adversarial learning [PITH_FULL_IMAGE:figures/full_fig_p002_4.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

12 extracted references

  1. [1]

    J. N. C. on Prevention, T. of High Blood Pressure, and N. H. B. P. E. Program,Report of the Joint National Committee on Prevention, Detection, Evaluation, and Treatment of High Blood Pressure, vol. 6. Public Health Service, National Institutes of Health, National Heart, Lung . . . , 1997

    1997

  2. [2]

    Accuracy of cuff-measured blood pressure: systematic reviews and meta-analyses,

    D. S. Picone, M. G. Schultz, P. Otahal, S. Aakhus, A. M. Al-Jumaily, J. A. Black, W. J. Bos, J. B. Chambers, C.-H. Chen, H.-M. Cheng, et al., “Accuracy of cuff-measured blood pressure: systematic reviews and meta-analyses,”Journal of the American College of Cardiology, vol. 70, no. 5, pp. 572–586, 2017

    2017

  3. [3]

    Pulse transit time based continuous cuffless blood pressure estimation: A new extension and a comprehensive evaluation,

    X. Ding, B. P. Yan, Y .-T. Zhang, J. Liu, N. Zhao, and H. K. Tsang, “Pulse transit time based continuous cuffless blood pressure estimation: A new extension and a comprehensive evaluation,”Scientific reports, vol. 7, no. 1, p. 11554, 2017

    2017

  4. [4]

    A multi- stage deep neural network model for blood pressure estimation using photoplethysmogram signals,

    J. Esmaelpoor, M. H. Moradi, and A. Kadkhodamohammadi, “A multi- stage deep neural network model for blood pressure estimation using photoplethysmogram signals,”Computers in Biology and Medicine, vol. 120, p. 103719, 2020

    2020

  5. [5]

    Investigation of data leakage in deep-learning-based blood pressure estimation using pho- toplethysmogram/electrocardiogram,

    R. Yoshizawa, K. Yamamoto, and T. Ohtsuki, “Investigation of data leakage in deep-learning-based blood pressure estimation using pho- toplethysmogram/electrocardiogram,”IEEE Sensors Journal, vol. 23, no. 12, pp. 13311–13318, 2023

    2023

  6. [6]

    Acnn-bilstm: A deep learning approach for continuous noninvasive blood pressure measurement using multi-wavelength ppg fusion,

    M. Cui, X. Dong, Y . Zhuang, S. Li, S. Yin, Z. Chen, and Y . Liang, “Acnn-bilstm: A deep learning approach for continuous noninvasive blood pressure measurement using multi-wavelength ppg fusion,”Bio- engineering, vol. 11, no. 4, p. 306, 2024

    2024

  7. [7]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. March, and V . Lempitsky, “Domain-adversarial training of neural networks,”Journal of machine learning research, vol. 17, no. 59, pp. 1–35, 2016

    2016

  8. [8]

    Automated curriculum learning for neural networks,

    A. Graves, M. G. Bellemare, J. Menick, R. Munos, and K. Kavukcuoglu, “Automated curriculum learning for neural networks,” ininternational conference on machine learning, pp. 1311–1320, Pmlr, 2017

    2017

  9. [9]

    Deep-learning-based blood pressure estimation using multi channel photoplethysmogram and finger pressure with attention mecha- nism,

    J. Kyung, J.-Y . Yang, J.-H. Choi, J.-H. Chang, S. Bae, J. Choi, and Y . Kim, “Deep-learning-based blood pressure estimation using multi channel photoplethysmogram and finger pressure with attention mecha- nism,”Scientific Reports, vol. 13, no. 1, p. 9311, 2023

    2023

  10. [10]

    Modfinity: Unsupervised domain adaptation with multimodal information flow intertwining,

    S. Liu, J. Lv, J. Kang, H. Zhang, Z. Liang, and S. He, “Modfinity: Unsupervised domain adaptation with multimodal information flow intertwining,” inProceedings of the Computer Vision and Pattern Recognition Conference, pp. 5092–5101, 2025

    2025

  11. [11]

    British hypertension society guidelines for hypertension management 2004 (bhs-iv): summary,

    B. Williams, N. R. Poulter, M. J. Brown, M. Davis, G. T. McInnes, J. F. Potter, P. S. Sever, and S. M. Thom, “British hypertension society guidelines for hypertension management 2004 (bhs-iv): summary,”Bmj, vol. 328, no. 7440, pp. 634–640, 2004

    2004

  12. [12]

    Ai-based multi-wavelength ppg device for blood pressure monitoring,

    C. Botrugno, K. Dheman, P. Bonazzi, F. Dell’Olio, and M. Magno, “Ai-based multi-wavelength ppg device for blood pressure monitoring,” in2024 IEEE International Symposium on Medical Measurements and Applications (MeMeA), pp. 1–6, IEEE, 2024

    2024

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.