acceptodds
Desk rejected at ICLR 2027

Learning What to Trust: Reliability-Conditioned Masked Modeling for Physiological Time Series

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Masked time-series models usually choose reconstruction targets without considering whether the observations are trustworthy. This is a poor fit for wearable physiological signals, where motion and contact artifacts create short, source-specific failures inside otherwise usable ECG and PPG recordings. We introduce QRAFT, a self-supervised learner in which fine-grained reliability determines which patches remain visible and which become reconstruction targets. Visible tokens are processed by a shared temporal encoder, allowing reliable evidence from either source to support local recovery in the other. A temporal, spectral, and contrastive objective trains the representation. To distinguish reliability-aware localization from generic hard-example mining, we introduce difficulty-matched random masking. At the same 38.5% mask ratio, QRAFT reduces PPG reconstruction RMSE by 23.2% relative to random masking and by 10.0% relative to the difficulty-matched control. The representation improves blood-pressure transfer across three datasets. On synchronized ECG–PPG data, fusion gains rise from 5.5% when both streams are clean to 18.6% when both are degraded. A factorial analysis further separates gains due to pretraining from those due to the fusion rule. These results show that local reliability can serve as supervision for masked representation learning, not merely as a filter applied before training

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
physiological time series, reliability-conditioned masking, self-supervised learning, masked representation learning, multimodal fusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_P1yQOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes QRAFT, a self-supervised framework for physiological time-series learning that uses local reliability estimates to determine which ECG and PPG patches should enter the encoder and which should become reconstruction targets. The method uses a weakly supervised quality estimator to generate patch-level reliability scores, followed by a shared encoder that combines information from ECG and PPG. It also uses temporal, spectral, and contrastive objectives for representation learning. The authors evaluate the method through reconstruction experiments, blood-pressure estimation, cross-dataset transfer, and quality-stratified comparisons. The experiments also include difficulty-matched random masking and factorial analysis to examine the contribution of reliability-conditioned masking and multimodal fusion.

Critical strengths
  • The paper proposes a clear local evidence selection idea. It uses a patch-level reliability proxy to determine which ECG and PPG patches may enter the encoder and which become reconstruction targets. I think this is a good contribution because the method explicitly considers local signal quality during representation learning.

  • Although there is a weakness regarding the use of coarse expert labels, as discussed below, the authors also report inter-rater agreement for those labels. This is a useful disclosure of annotation reliability and the uncertainty associated with weak supervision.

  • The experimental section is quite comprehensive. The authors compare the method with several baselines and provide clear quality-stratified blood-pressure performance results. They also include difficulty-matched masking and factorial experiments to examine the contributions of different components.

  • It is also nice to see that removing quality perception leads to worse performance in Table 3. This is consistent with the hypothesis that local quality information contributes to the model's performance beyond simply increasing model capacity.

Critical weaknesses
  • There is a little bit of overclaiming regarding patch-level reliability. The reliability estimator is not directly supervised with patch-level artifact annotations. Instead, the reliability scores are learned from coarse record-level expert labels and then used to select patches for masking. Therefore, the reported quality-estimation accuracy does not directly establish that the selected patches coincide with actual artifact intervals. I think it would be better if the authors provided patch-level annotations or a controlled corruption experiment with known artifact locations to evaluate the reliability of the selected patches. I understand that coarse labels may be the best available supervision during training, but from an evaluation perspective, a more direct ground truth would make the results more convincing. The difficulty-matched masking and factorial comparisons support the claim that the proposed selection policy contributes to performance gains. However, these experiments do not establish whether the gains come specifically from accurate artifact localization or from other properties of the selection policy. The paper already evaluates robustness under synthetic corruption, but that is different from directly measuring patch-level localization accuracy.

  • There is a potential problem around lines 209–211. The method masks low-reliability patches but computes reconstruction loss against the recorded waveform, which may contain the artifact that triggered the masking. The authors explicitly state that the reconstruction targets include any artifacts present in the original observations. Therefore, lower reconstruction error against the recorded patches does not necessarily mean that the model recovers the underlying clean ECG or PPG signal. There are several things the model could be learning through this loss, and it is difficult to distinguish waveform prediction from physiological restoration. It would be better if the authors included a controlled corruption experiment with a known clean signal, so that reconstruction performance could be evaluated separately from the recovery of physiologically correct waveforms.

  • Figure 2 is really difficult to read. It contains four panels, and each panel is extremely dense. The architecture and the relationships between components are hard to follow. I think the authors could improve the figure by simplifying the diagrams or separating some of the components. The presentation could be polished further to make the method easier to understand.

  • Regarding the experimental setup, the authors use 1,000 subjects from VitalDB for self-supervised pretraining. I would have expected a larger and more diverse subject pool for a general physiological representation-learning framework. However, the number of subjects alone does not establish that the dataset is insufficient, especially because each subject can provide multiple recording segments. My concern is whether the reported performance and generalization would remain stable with a larger and more diverse pretraining population. A dataset-size ablation or additional pretraining data could help clarify this.

  • I do not fully understand the purpose of Equation (9). The authors explicitly state that it does not introduce a mutual-information term into the training objective. Therefore, the equation appears to provide a theoretical interpretation rather than something directly optimized by the model. I understand that an interpretation can still be useful, but its relevance to the proposed method is not explained clearly enough. The authors could explain more directly how this information-theoretic bound motivates the reliability-conditioned masking strategy and what additional insight it provides beyond the existing reconstruction objective.

Questions for authors

The following questions refer to the weaknesses discussed above.

  • Could the authors clarify how accurately the learned patch-level reliability scores identify actual artifact intervals? Is there evidence from patch-level annotations or controlled corruption with known artifact locations?

  • Given that the reconstruction targets may contain artifacts, how should the reported reconstruction improvements be interpreted? Is there evidence that the model can recover physiologically correct signals rather than simply predict the recorded waveforms?

  • Could the authors clarify whether the performance gains remain stable with a larger or more diverse pretraining population? Is there any evidence of how performance changes with the number of pretraining subjects?

4 comments
AuthorsOpenReview ↗

Response to Reviewer P1yQ (Part 1/2: W1–W4)

Thank you for the constructive comments on patch reliability, reconstruction targets, pretraining scale, and presentation. We address W1–W5 first, followed by Q1–Q3.


W1: Record-level quality supervision and QA accuracy do not establish patch-level artifact localization. Please provide patch annotations or controlled corruption with known locations; difficulty-matched and factorial comparisons address a different question.

A1: Experts provide record-level quality labels; FGQA learns half-second instance scores through multi-instance learning. We therefore evaluate localization directly using known synthetic artifact locations, separately from QA classification and downstream robustness.

In each held-out 10-second PPG segment, we inject motion artifacts into 4 of 20 half-second patches. Across 20,000 test patches, including 4,000 positives, FGQA achieves AUPRC 0.80. Positive prevalence is 0.20, the random-ranking reference. This controlled evaluation directly tests whether the learned scores identify the injected regions.

Revision locations.

  • Supplement §B.1 | p. 7 — Synthetic artifact localization protocol, test size, AUPRC, and positive prevalence.

W2: Reconstruction targets include artifacts in the recorded waveforms. Lower reconstruction error does not establish clean physiological recovery. Please evaluate prediction of recorded observations separately from recovery against a known clean signal.

A2: The policy-specific reconstruction evaluation (original Table 2) uses recorded PPG as the target. In the shared-mask reconstruction evaluation, all methods predict identical randomly withheld patches at a 38.5% ratio, with overall RMSE of 0.72 for QRAFT versus 0.82 for Random mask. In the paired clean-reference evaluation, we score regions with controlled synthetic motion artifacts against their pre-injection PPG. With the Random-mask clean-reference RMSE normalized to 1.00, QRAFT reaches 0.91, a 9% reduction. The clean-reference score measures recovery under the specified corruption protocol and complements the recorded-waveform evaluation.

Revision locations.

  • Supplement §§C.1–C.3 | p. 12 — Policy-specific, shared-mask, and paired clean-reference evaluations.

W3: Figure 2 is dense, and the relationships between its four panels are difficult to follow. Please simplify the diagrams or separate components to improve readability.

A3: Figure 2 now presents quality scoring, patch selection, shared encoding, and source-specific reconstruction as a single workflow. It makes the role of reliability explicit: selected patches provide visible context, and source-specific decoders predict the withheld waveform content. Supplementary Figure S1 provides the detailed module diagram.

Revision locations.

  • Main text Figure 2 | p. 5 — Quality-selection, encoding, and reconstruction workflow.
  • Supplement §A.2 | p. 4, Figure S1 — Detailed module diagram.

W4: Pretraining uses 1,000 VitalDB subjects. Would performance remain stable with a larger or more diverse population? A dataset-size ablation or additional pretraining data would clarify scale sensitivity.

A4: We vary the fraction of the selected 1,000-subject VitalDB pretraining pool while keeping the labeled MIMIC-III fine-tuning subset fixed.

R3-Table1: MIMIC-III SBP MAE by fraction of the VitalDB pretraining pool.

Fraction of pretraining pool1%5%10%25%50%75%100%
MIMIC-III SBP MAE (mmHg)8.836.975.824.964.374.104.09

SBP MAE decreases as the VitalDB pool grows, reaching 4.10 mmHg at 75% and 4.09 mmHg at 100%. The sweep shows diminishing gains as the pool approaches 1,000 subjects and evaluates pretraining size within this cohort.

Revision locations.

  • Supplement §E | p. 19, Table S34 — VitalDB scale sweep with a fixed MIMIC-III fine-tuning subset.

AuthorsOpenReview ↗

Response to Reviewer P1yQ (Part 2/2: W5, Q1–Q3)

W5: Equation (9) is an information-theoretic interpretation rather than an optimized mutual-information term. How does it motivate reliability-conditioned masking, and what insight does it add beyond reconstruction?

A5: For a fixed mask policy and fixed-variance Gaussian decoder, reducing squared reconstruction error raises a variational lower bound on information retained about withheld observations. Reliability-conditioned masking chooses both the visible evidence and the prediction targets: reliable patches supply context for predicting withheld waveform content.

The insight from Equation (9) is that mask selection determines both the evidence available to the encoder and the content the representation is trained to predict. Equation (8) defines the training objective. Table 2 and the shared-mask evaluation test the masking strategy empirically.

Revision locations.

  • Main text §3.3 | p. 5 — Information-theoretic interpretation of reconstruction.
  • Supplement §C.2 | p. 12, Table S18 — Empirical comparison on shared reconstruction targets.

Q1: Could the authors clarify how accurately the learned patch-level reliability scores identify actual artifact intervals? Is there evidence from patch-level annotations or controlled corruption with known artifact locations?

A6: The known-location experiment directly evaluates the learned patch scores. Each held-out 10-second PPG segment contains 20 half-second patches, with synthetic motion artifacts injected into four. FGQA achieves AUPRC 0.80 across 20,000 patches, including 4,000 injected positives; the random-ranking reference is 0.20. These results assess localization of controlled synthetic artifacts, complementing the record-level expert supervision used to train FGQA.

Revision locations.

  • Supplement §B.1 | p. 7 — Patch-level localization evaluation and results.

Q2: Given that the reconstruction targets may contain artifacts, how should the reported reconstruction improvements be interpreted? Is there evidence that the model can recover physiologically correct signals rather than simply predict the recorded waveforms?

A7: The recorded-waveform results measure prediction of observed PPG. To evaluate recovery separately, we use paired pre-injection PPG as the reference and score predictions within the injected artifact regions.

R3-Table2: Paired clean-reference RMSE within synthetically corrupted regions, normalized by the Random-mask result.

Pretraining methodRelative clean-reference RMSEReduction versus Random mask
Random mask1.00Reference
QRAFT0.919%

The 9% reduction supports recovery toward the pre-injection waveform under this controlled protocol. The separate shared-mask comparison, 0.72 versus 0.82 RMSE, evaluates prediction of identical recorded-waveform targets.

Revision locations.

  • Supplement §§C.2–C.3 | p. 12 — Shared-mask comparison and paired clean-reference recovery.

Q3: Could the authors clarify whether the performance gains remain stable with a larger or more diverse pretraining population? Is there any evidence of how performance changes with the number of pretraining subjects?

A8: R3-Table1 reports the available scale evidence with a fixed MIMIC-III fine-tuning subset. SBP MAE falls from 8.83 to 4.96 mmHg as the VitalDB fraction increases from 1% to 25%, and from 4.10 to 4.09 mmHg between 75% and 100%. Thus, increasing the pool improves downstream performance, with smaller changes near the full 1,000-subject pool. This experiment evaluates pretraining size within VitalDB.

Revision locations.

  • Supplement §E | p. 19, Table S34 — Performance across seven fractions of the VitalDB pretraining pool.

The revisions clarify Equation (9) and provide controlled evidence for artifact localization, clean-reference recovery, and pretraining-scale effects. We thank Reviewer P1yQ for the constructive suggestions and welcome further questions.

Thank you for running the controlled experiments I asked for. The synthetic artifact-localization result (AUPRC 0.80 vs 0.20 for random ranking) and the clean-reference recovery evaluation (9% improvement over random masking) address my two main questions directly, and the dataset-size ablation is useful. My remaining concerns are narrower:

  1. Both new experiments are synthetic, so patch-level accuracy on real annotated artifacts is still open; and
  2. The scale ablation varies the fraction of the same 1,000-subject cohort, so generalization to a larger and more diverse population is untested

I am keeping my score at 3 (weak accept) because I feel like a straight 4 is not up to the quality of this paper, but maybe in my mind the score is up to around 3.4.

AuthorsOpenReview ↗

Response to Reviewer P1yQ: Real-signal validation and population extension

Thank you for confirming that the controlled localization and clean-reference experiments directly address your two main questions, and for recognizing the value of the size ablation. Your remaining concerns distinguish natural-artifact localization from local quality assessment, and broader-population pretraining from downstream transfer. We provide additional completed evidence and an update on the population extension below.


1. Validation on real artifacts beyond synthetic corruption.

A1: Real-signal expert-score validation. We additionally compare the current FGQA scores with expert assessments on real ECG and PPG. Three experts scored nine segments per modality, with 20 scored positions per segment on a 0–100 scale.

Follow-up Table 1: Expert absolute agreement and FGQA correlation with mean expert scores.

ModalityExpert ICC(A,1) [95% CI]FGQA–expert correlation
ECG0.871 [0.694, 0.950]0.94
PPG0.939 [0.919, 0.954]0.97

ICC confidence intervals use 10,000 whole-segment bootstrap resamples. On this small real-recording subset, the correlations provide evidence of agreement between FGQA and expert-rated local signal quality. This complements the known-location synthetic experiment; localization against annotated natural-artifact intervals remains a separate evaluation.

Physiological-feature preservation. The existing reconstruction analysis also measures identifiable waveform features:

Follow-up Table 2: Physiological-feature preservation under reconstruction.

MetricRandom-maskQRAFT
Systolic peak retention (%) ↑91.7 ± 0.7895.2 ± 0.28
Diastolic peak retention (%) ↑89.5 ± 1.0394.1 ± 0.37
Heart-rate error (bpm) ↓2.42 ± 0.081.35 ± 0.05

The peak-retention and heart-rate measurements add physiological-feature evidence to the paired clean-reference recovery result.

Revision locations.

  • Supplement §B.2 | p. 6, Table S9 — Expert-score agreement and FGQA validation.
  • Supplement §C | p. 7, Table S13 — Physiological-feature preservation.

2. Larger and more diverse pretraining populations.

A2: The reported size ablation characterizes scaling within the selected 1,000-subject VitalDB cohort. Evaluating pretraining on a larger and more diverse population is the next stage of this work.

Additional completed transfer experiment. We further evaluate transfer of the pretrained encoders to heartbeat classification on the MIT-BIH Arrhythmia Database. Each encoder is frozen, and a linear classifier predicts the five AAMI classes (N, S, V, F, Q).

Follow-up Table 3: MIT-BIH frozen-encoder classification; mean ± SD over ten paired seeds.

Pretraining methodAccuracy (%) ↑Macro-F1 (%) ↑
Random-mask reconstruction94.2 ± 0.885.1 ± 1.2
DM-Rand94.8 ± 0.786.0 ± 1.1
CPC95.1 ± 0.786.6 ± 1.0
JEPA95.4 ± 0.687.2 ± 0.9
QRAFT95.9 ± 0.588.1 ± 0.8

Methods share the pretraining data and matched training budgets. The fixed patient-independent split contains 22 training records and 21 test records. Training uses class weighting; testing retains the original class distribution.

QRAFT improves Accuracy by 1.7 percentage points and Macro-F1 by 3.0 percentage points over random-mask pretraining. This provides complementary evidence of heartbeat-class information accessible through a linear classifier on a separate real ECG dataset.

The original BP-transfer experiments evaluate the VitalDB-pretrained model on MIMIC-III, UCI, and the in-house cohort (IHC); the MIT-BIH experiment adds a different downstream task, providing complementary transfer evidence across datasets and tasks.

Population-extension progress. As a continuation of this work, we have collected public physiological data covering more than 10,000 participants from MIMIC, VitalDB, MESA, UCI, and other resources, alongside our in-house multicenter database. Pretraining on this expanded collection requires a longer training cycle, and preliminary results are not yet available.

Revision locations.

  • Supplement §E.2 | p. 18, Tables S34–S35 — MIT-BIH classification results and patient-independent split.
  • Supplement §E.3 | p. 21, Table S41 — VitalDB pretraining-size ablation.

Thank you for recognizing our work and the additional evidence. Whatever the final evaluation outcome, your thoughtful comments have helped us substantially improve the experimental evaluation and explain the findings more clearly. We appreciate the time and care you have devoted to this discussion, and wish you continued success in your research.

Reviewer_YTEVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In "Learning What to Trust: Reliability-Conditioned Masked Modeling for Physiological Time Series", the authors investigate the impact of the masking strategy on masked reconstruction-type self-supervised representation learning for physiological time series data (ECG and PPG data). The authors argue for a reliability-aware masking, where reliable patches remain visible and unreliable patches are masked and used as prediction targets. Evaluation metrics are mostly focused on reconstruction quality.

Critical strengths
  • Investigating the inclusion of signal quality measures into self-supervised representation learning is a worthwhile endeavor as poor signal quality segments are a central challenge when dealing with these datasets.
Critical weaknesses

Major

  • General suitability of masked reconstruction: It is known that features for reconstruction and features for semantic tasks do not align (at least in vision https://arxiv.org/abs/2402.11337). This puts some doubts on the suitability of masked reconstruction tasks also as a pretraining objective for physiological data.
  • Metrics: The paper uses reconstruction metrics as their primary metric, which seems inappropriate in the light of the discrepancy between reconstruction and semantic quality as mentioned above. Secondary metrics include BP prediction from PPG, which is a very challenging task that is essentially unsolved (in the sense of reaching clinical grade performance) on large-scale cohorts without excessive quality filtering and therefore not really reliable as downstream evaluation task. Table 4 mentions further downstream tasks that are not properly introduced neither in the main text nor in the appendix.
  • Random-masking: I do not see clear evidence that difficulty-matched random masking is possible based on local variance, sample entropy, and high-frequency energy. Eventually, segments identified via signal quality vs. difficulty-matched random sampling will showing very different characteristics (e.g. when training a classifier on them) that go beyond the three observables mentioned above.
  • Missing technical details: The description of the FGQA (Section 3.1) is very hard to parse. The data basis underlying the training remains unclear to me. Are the training data i.e. annotations publicly available/provided? Similarly, Section 3.2 is in parts very schematic and does not specify any details about stochastic depth, nature of the reconstruction heads. In Section 3.3, the augmentations used for the contrastive loss remain unspecified. The authors specify very specific hyperparameters for weighting the loss terms below equation (8), whose origin is not dicussed.
  • ECG data in MIMIC-III: ECG data in MIMIC-III is compromized (see Physionet's readme page), which essentially makes it unusable for deep learning analyses. There is no mentioning in the manuscript on how the authors circumvent this challenge.
  • Discussion: The discussion is very limited and recapitulates only already identified points. In particular, it does not discuss any limitations of the approach.

Minor

  • According to my understanding the symbol U+22AE in should indicate an indicator function, this is a non-standard notation introduced nowhere.
  • The abbreviation FGQA is never introduced.

Summary

  • The general suitability of the masked reconstruction task remains doubtful in particular in the light of reconstruction metrics as primary evaluation metrics with unsuitable or not properly defined other target metrics. The paper is partially hard to parse and misses essential technical details.
Questions for authors

Questions

  • Do the authors have any thoughts on the prospects of masked reconstruction objectives for physiological time series in general? Judging from foundation models trained for ECG/PPG, it seems that rather latent prediction models (JEPA-style or CPC) are the preferred approaches.
  • Could the authors clarify which masking strategy was used for the evaluation in Table 1/2. Anything apart from random masking would provide an unfair advantage for methods that were trained using a similar masking strategy.
  • Could the authors provide an intuitive argument for the strategy "reliable patches remain visible and unreliable patches are masked and used as prediction targets"? (a) During inference time the model will have to infer representations for any segment, even in low signal quality environments. (b) Conditioning on reliable patches seems an appropriate strategy to to provide appropriate conditioning. It seems to me that the downside of this is that using low-quality segments as prediction targets limits the reliability of the prediction targets, which could compromize the quality of the learned representation. In any case, the strategy does not seem as clear-cut as the author sell it in the introduction without further justification.
12 comments
AuthorsOpenReview ↗

Response to Reviewer YTEV (Part 1/4: W1–W3)

W1: The suitability of masked reconstruction for physiological representation learning is questioned.

A1: QRAFT studies reliability-conditioned evidence selection within masked modeling. Waveform prediction supplies temporal and cross-source structure, while BP transfer and auxiliary tasks evaluate downstream representation utility. On held-out MIMIC-III subjects with full PPG input, the QRAFT-pretrained checkpoint reaches 4.20 mmHg SBP MAE versus 4.58 mmHg for the random-mask checkpoint. With FGQA gating applied to both fixed checkpoints, the corresponding MAEs are 4.09 and 4.50 mmHg. Auxiliary tasks include epilepsy classification, activity recognition, sleep staging, and ECG/PPG quality classification (Main text §4.1; Supplement §E).


W2: Reconstruction metrics and BP prediction require complementary downstream evidence.

A2: We separate target prediction from representation transfer. Policy-specific reconstruction, shared-mask reconstruction, fixed-checkpoint BP transfer, linear probing, fine-tuning, and auxiliary classification use distinct evaluation protocols. The shared-mask comparison gives 0.72 RMSE for QRAFT, 0.77 for DM-Rand, and 0.82 for Random mask; the fixed-checkpoint comparison gives 4.09 mmHg for QRAFT with FGQA gating and 4.20 mmHg with full PPG (Main text §§4.1–4.2; Supplement §§C.2, D.1).


W3: The fairness of difficulty-matched random masking requires a controlled comparison.

A3: DM-Rand selects difficult targets using normalized local variance, normalized sample entropy, and log energy ratio above 15 Hz. All methods use the same 38.5% mask ratio. In the shared-mask evaluation, every method predicts the same PPG test targets, with QRAFT at 0.72 RMSE versus 0.82 for Random mask and 0.77 for DM-Rand (Main text §4.2; Supplement §C.2).


AuthorsOpenReview ↗

Response to Reviewer YTEV (Part 2/4: W4–W8)

W4: FGQA, the architecture, augmentations, loss weights, and notation require explicit definitions.

A4: FGQA uses PPG resources totaling 71,800 five-second fragments: MIMIC-III 18,000, UCI 12,000, and Queensland 41,800. The ECG resource contains 10,918 PhysioNet/CinC 2017 segments, 15,659 BUTQDB segments, and 10,000 NSTDB-noise augmented samples. Three trained clinical-signal experts annotated the complete ECG and PPG resources; two matching judgments were retained, and disagreements were adjudicated by the third. The labels are record-level, and FGQA learns 0.5-second scores through multi-instance learning. The annotation set overlaps BP training/validation subjects and is disjoint from the final reconstruction and BP-transfer test subjects; the FGQA threshold is calibrated on a separate held-out split. The public release contains model and preprocessing code, while the source annotation files are not included. The implementation specifies stochastic depth 0–0.10, two-layer 128-dimensional source-specific reconstruction heads, symmetric InfoNCE with temperature 0.07, four augmentations, temporal weights 0.75/0.25, and contrastive/spectral weights 0.20/0.10. Supplement C reports the corresponding weight sensitivity sweep, with PPG RMSE ranging from 0.61 to 0.70 across settings. Main text §§3.1–3.3; Supplement §§A.2–C give the complete definitions.



W5: MIMIC-III ECG quality and ECG–PPG preprocessing require an explicit protocol.

A5: Table 5 uses ECG lead II and PPG/PLETH from the same waveform segment. The preprocessing excludes 21,400 records (18.5%) with missing or unusable waveforms, uses acquisition timestamps for alignment, resamples both signals to 125 Hz, and samples 400 ten-second segments per subject. Table 1 uses PPG only, whereas Table 5 uses paired ECG–PPG input; pretraining uses the separate VitalDB cohort. The protocol is timestamp-aligned ECG–PPG fusion under the documented MIMIC-III waveform setting (PhysioNet MIMIC-III Waveform Database).



W6: The Discussion should explain how ECG sampling and temporal correspondence affect fusion.

A6: The revised Discussion connects sampling jitter and changing inter-waveform delay to multimodal temporal correspondence. These factors affect how ECG and PPG evidence can support one another after timestamp alignment, especially when only a small subset of local patches remains reliable. The Discussion links this preprocessing setting to the observed fusion behavior and identifies the role of remaining temporal and cross-source context (Main text Discussion and Limitations, p. 9).



W7: The indicator notation is undefined.

A7: In Eq. (1), 1[·] denotes the binary indicator function: it equals 1 when the condition inside the brackets is satisfied and 0 otherwise. The indicator assigns each patch to the visible evidence set or the masked reconstruction-target set according to its FGQA score.



W8: FGQA is not expanded at first use.

A8: FGQA denotes Fine-Grained Quality Assessment. It is the local quality estimator that receives record-level supervision, learns 0.5-second instance scores through multi-instance learning, and supplies the scores used to construct the masking policy. The related QA terminology is defined in Supplement §A.2.



AuthorsOpenReview ↗

Response to Reviewer YTEV (Part 3/4: Q1)

Q1: Do the authors have any thoughts on the prospects of masked reconstruction objectives for physiological time series in general? Judging from foundation models trained for ECG/PPG, it seems that rather latent prediction models (JEPA-style or CPC) are the preferred approaches. The paper uses reconstruction metrics as their primary metric, which seems inappropriate in the light of the discrepancy between reconstruction and semantic quality as mentioned above. Secondary metrics include BP prediction from PPG, which is a very challenging task that is essentially unsolved (in the sense of reaching clinical grade performance) on large-scale cohorts without excessive quality filtering and therefore not really reliable as downstream evaluation task. Table 4 mentions further downstream tasks that are not properly introduced neither in the main text nor in the appendix.

A9: QRAFT studies reliability-conditioned evidence selection for masked modeling. Waveform prediction supplies a training signal for temporal and cross-source structure, while frozen linear probes and fine-tuning assess downstream transfer. JEPA and CPC use latent prediction; QRAFT uses local reliability to define the evidence and reconstruction targets. On held-out MIMIC-III subjects with full PPG input, the QRAFT-pretrained checkpoint reaches 4.20 mmHg SBP MAE versus 4.58 mmHg for the random-mask checkpoint. With FGQA gating applied to both fixed checkpoints, the corresponding MAEs are 4.09 and 4.50 mmHg. The inference-gating ablation separates checkpoint and gating effects.

The BP benchmark uses subject-level 6:2:2 training/validation/test splits and PPG inputs, with no subject-specific BP calibration or offset. R2-Table2 reports both checkpoints with full PPG and with FGQA gating.

Table 4 covers epilepsy classification, human activity recognition, sleep staging, and ECG and PPG signal-quality classification, evaluated by accuracy and F1 under linear probing and fine-tuning. These tasks use random-mask pretraining and assess encoder transfer. Supplementary Section E now gives the task definitions and full classification results.

Revision locations.

  • Main text §4.1 | p. 6 — Subject-level BP evaluation protocol.
  • Supplement §D.1 | p. 16, Table S25 — Matched-input BP comparisons.
  • Supplement §E | pp. 17–18, Table S29 — Auxiliary task definitions and encoder-transfer results.

Reviewer_F4NCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

QRAFT is a masked self-supervised learner for synchronized ECG and PPG in which a weakly supervised local quality estimator (FGQA, trained by multi-instance learning from record-level expert labels) decides which patches stay visible and which become reconstruction targets: low-quality patches are masked. Visible ECG and PPG tokens are encoded in one shared transformer so that reliable context in one source can support reconstruction of withheld patches in the other, with source-specific decoders and a time-domain, log-spectral and contrastive objective. A difficulty-matched random masking control (DM-Rand) and a factorial over pretraining policy and fusion rule are introduced to separate the contributions. At a matched mask ratio QRAFT reduces PPG reconstruction RMSE from 0.82 (random) and 0.70 (DM-Rand) to 0.63 (Table 2), obtains the lowest blood-pressure errors on MIMIC-III, UCI and an in-house cohort (SBP MAE 4.09 mmHg on MIMIC-III after fine-tuning, Table 1), and in synchronized ECG-PPG fusion reaches SBP MAE 3.36 mmHg versus 3.57 for soft quality weighting (Table 5), with fusion gains growing from 5.5% when both sources are clean to 18.6% when both are degraded (Section 4.5).

Critical strengths

1. The core idea is simple, well motivated, and tested with the right controls.

Using local signal quality to decide what counts as evidence and what the model must predict, rather than filtering records or weighting modalities after encoding, is a natural fit for wearable signals.

The difficulty-matched control (DM-Rand, Section 4.2) and the pretraining-by-fusion factorial (Section 4.3) are the right experiments to separate this idea from hard-example mining and from the fusion rule, and the quality-regime analysis (Section 4.5) behaves as the hypothesis predicts.

2. Gains are consistent across settings and the work is reproducible in principle.

Improvements appear in reconstruction (Table 2, ten seeds, paired Wilcoxon), cross-dataset BP transfer on three cohorts (Table 1), and synchronized fusion (Table 5). Code is provided.

Critical weaknesses

I recommend rejection for the two reasons below: the manuscript cannot be assessed or reproduced as written because key components are undefined, and the evidence does not isolate the titular contribution while the headline numbers lack the protocol detail needed to believe them. I state in the Questions what would change my recommendation.

1. The manuscript is not self-contained, and the component on which everything rests is undocumented.

The contrastive term L_c in Equation 8 is never defined (no formula, augmentations or temperature). The "QRAFT fuser" that Section 4.3 credits with the larger share of the gain is not defined anywhere, and Comparisons I to III in Table 3 are not defined in the main text; I looked for both in the supplement and found only names, not definitions. Symbols and acronyms are used without definition (e_i^r, t_mask, p_i, g_psi in Equation 1; FGQA, QA, AC Encoder, MSFA, MHC, ECA, Mix-T, TSR-Loss, PPG-PM, PI).

Most importantly, the provenance of the coarse expert quality labels that train FGQA is never stated: which cohort, how many 5-second records, which annotators, and whether this labeled set overlaps the reconstruction or transfer test data. The masking policy, and hence the whole method, is derived from these labels. As written, Section 3 cannot be reproduced or assessed from the paper.

2. The evidence does not isolate the titular contribution, and the headline BP numbers need protocol detail before they can be believed.

In the factorial decomposition reported in Section 4.3, the fusion rule contributes 5.8% to 7.9% while reliability-conditioned pretraining contributes less (3.4% to 5.6% in the supplement's full table). The masking gain is confined to low-quality patches (RMSE 0.92 vs 1.08 for DM-Rand) and disappears on high-quality ones (0.41 vs 0.40), both from the subset rows of Table 2.

Because masked targets are the observed patches including their artifacts (Section 3.3), the improvement is a better reconstruction of artifacts. The paper does not argue why this is a desirable pretext objective, nor compare with the reverse policy (mask reliable, keep unreliable visible) or with excluding unreliable patches from both evidence and targets.

A calibration-free, subject-independent SBP MAE of 4.09 mmHg on MIMIC-III (Table 1) is far below what is typically reported for subject-independent PPG-based BP estimation. The main text gives no split protocol for MIMIC-III or UCI; because this number decides the paper's credibility I checked the supplement, which states only "1,000 subjects and 400 records per subject" and an error SD of 5.20 mmHg, which raises rather than settles the question. The split unit, test-subject disjointness and any per-subject calibration must be spelled out in the main text. Table 1 is five-run descriptive without confidence intervals, and Table 4 uses random masking, so it does not test the proposed method.

Questions for authors

Q1 to Q3 would materially change my assessment.

  1. Where do the coarse expert labels come from (dataset, number of 5-second records per modality, annotator background), and is the labeled set disjoint from the reconstruction test split and from the BP transfer cohorts?

  2. MIMIC-III and UCI protocol: how were subjects split; are all windows of a test subject unseen during pretraining and fine-tuning; was any per-subject calibration or offset used; how were the 400 records per subject (Supplementary Section A) sampled; and which input modalities feed Table 1 (PPG only or ECG plus PPG)? Please add confidence intervals to Table 1.

  3. Please define L_c (formula, augmentations, temperature), the QRAFT fuser versus cross-attention fusion, and Comparisons I to III in Table 3, and expand FGQA and the other acronyms at first use.

  4. Why is reconstructing artifact-containing patches a good pretext target? Please report the reverse policy (mask reliable patches, keep unreliable visible) and the "exclude unreliable from both evidence and targets" policy at the same 38.5% ratio used in Section 4.2.

  5. Since the masking gain vanishes on high-quality signals, what is the expected benefit in cohorts where most data are clean? A plot of the BP gain against the fraction of low-quality patches would answer this with existing predictions.

  6. Suggestions not affecting the score: Table 4 (random-mask pretraining on unrelated classification tasks) tests the backbone rather than the method and could move to the supplement. Please also state the ethics approval number for the in-house cohort (the institution can remain anonymized) and whether the in-house acquisition device described in the supplement required any additional approval for use on human participants.

5 comments
AuthorsOpenReview ↗

Response to Reviewer F4NC (Part 1/4: W1–W2, Q1)

W1: The manuscript is not self-contained: key method definitions, acronyms, and the provenance of the FGQA supervision are missing.

A-W1: FGQA supplies record-level quality supervision, and its multi-instance learner converts these labels into 0.5-second instance scores used by the masking policy. The PPG resource contains 71,800 five-second fragments from MIMIC-III, UCI, and Queensland. The ECG resource contains 10,918 PhysioNet/CinC 2017 segments, 15,659 BUTQDB segments, and 10,000 NSTDB-noise augmented samples. Three trained clinical physiological-signal experts annotated the complete ECG and PPG resources; two concordant judgments were retained and disagreements were adjudicated by the third.

QRAFT uses the resulting scores to define the visible evidence and reconstruction targets. The visible ECG/PPG tokens are jointly encoded, while source-specific heads reconstruct the masked observations. The revised manuscript gives the LcL_c formula, temperature and augmentations, QRAFT-FUSE, Comparisons I–III, symbols, and acronyms. The full specification is in Main text §§3.1–3.3 and Supplement §A.2.


W2: The evidence does not fully isolate the reliability-conditioned contribution, and the BP protocol needs explicit details.

A-W2: The 2×2 factorial separates pretraining and fusion: QRAFT pretraining reduces SBP MAE by 3.4%–5.6% with fusion fixed, while QRAFT-FUSE reduces it by 5.8%–7.9% with pretraining fixed. The masking controls provide the complementary comparison: at the matched 38.5% ratio, QRAFT gives 0.63 RMSE, compared with 0.70 for DM-Rand, 0.75 for Reverse mask, and 0.89 for Exclude-both. Under identical random test masks, QRAFT reaches 0.72 RMSE versus 0.82 for Random mask.

The BP protocol uses subject-level 6:2:2 splits, PPG input for Table 1, 400 randomly sampled ten-second MIMIC-III segments per subject, no subject-specific calibration or offset, and five-run mean ± SD with across-run intervals. Under fixed MIMIC-III checkpoints, FGQA gating changes QRAFT from 4.20 to 4.09 mmHg. Quality-stratified BP analysis and the in-house ethics information are reported below.


Q1: Where do the coarse expert labels come from (dataset, number of 5-second records per modality, annotator background), and is the labeled set disjoint from the reconstruction test split and from the BP transfer cohorts?

A1: The FGQA resources follow the source-study database construction. The PPG resource contains 71,800 five-second fragments: MIMIC-III (18,000), UCI (12,000), and Queensland (41,800). The ECG resource contains 10,918 processed segments from PhysioNet/CinC 2017 and 15,659 segments from BUTQDB; 10,000 additional poor-quality ECG samples were generated with MIT-BIH NSTDB noise.

PPG quality was annotated as Good, Acceptable, or Bad, with Good and Acceptable combined for the acceptable class. ECG quality used the source labels and the Good/Poor grouping described for 2017PNDB and BUTQDB. Three trained clinical physiological-signal experts annotated the complete FGQA ECG and PPG resources. Two concordant judgments were retained, and disagreements were adjudicated by the third. The expert annotations are record-level labels; FGQA learns 0.5-second instance scores through multi-instance learning.

The FGQA annotation set shares some subjects with the BP training and validation sets. The full annotation set is disjoint at the subject level from the final reconstruction and BP-transfer test sets. We calibrate the FGQA threshold on a separate held-out split.

Revision locations.

  • Main text §3.1 | pp. 3–4 — FGQA annotation sources and final-test subject separation.
  • Supplement §A.2 | p. 3 — Annotation protocol, sources, and overlap with BP development subjects.
AuthorsOpenReview ↗

Response to Reviewer F4NC (Part 2/4: Q2)

Q2: MIMIC-III and UCI protocol: how were subjects split; are all windows of a test subject unseen during pretraining and fine-tuning; was any per-subject calibration or offset used; how were the 400 records per subject (Supplementary Section A) sampled; and which input modalities feed Table 1 (PPG only or ECG plus PPG)? Please add confidence intervals to Table 1.

A4: MIMIC-III and UCI use subject-level training/validation/test splits of 6:2:2. MIMIC-III therefore uses 600/200/200 subjects for training/validation/test. UCI contains 942 patients under the same subject-level protocol; all windows from a subject remain in one partition. Pretraining uses 1,000 VitalDB subjects from a separate cohort. Downstream test subjects are excluded from pretraining and fine-tuning.

MIMIC-III contains 1,000 selected subjects, with 400 ten-second segments randomly sampled per subject. UCI contains 12,000 records from 942 patients. Signals are resampled to 125 Hz. Table 1 uses PPG only. QRAFT applies FGQA gating at inference, while the baselines receive the full PPG input. QA-filtered MAE screens low-quality records before pretraining. Table 5 uses paired ECG lead II and PPG/PLETH segments where both are available. We apply no subject-specific BP calibration or offset. Arterial BP supplies the downstream target.

Table 1 reports five-run mean ± SD. The intervals below summarize the corresponding five run-level results and are reported as approximate 95% across-run intervals:

R1-Table1: Fine-tuned SBP MAE and approximate 95% confidence intervals across five runs.

DatasetQRAFT fine-tuned SBP MAE (mmHg)Approx. 95% CI across five runs
MIMIC-III4.09 ± 0.17[3.88, 4.30]
UCI3.52 ± 0.14[3.35, 3.69]
In-house cohort4.55 ± 0.16[4.35, 4.75]

R1-Supplementary Table S1 gives the intervals for all Table 1 entries. These intervals describe run-to-run variation; they are not subject-level intervals. Supplement D compares hard, thresholded-soft, and soft gating, with SBP MAEs of 4.09, 4.17, and 4.31 mmHg, respectively.

Both checkpoints were fine-tuned on full PPG. On the original Table 1 MIMIC-III test set, we held each checkpoint fixed and toggled FGQA only at inference.

R1-Table2: Effect of inference gating on SBP MAE (mmHg).

BP checkpointFull PPGFGQA-gated PPGMAE reduction
Random-mask MAE4.584.500.08
QRAFT4.204.090.11

At matched inference inputs, QRAFT has a lower SBP MAE by 0.38 mmHg with full PPG and by 0.41 mmHg with FGQA gating.

Revision locations.

  • Main text §4.1 | p. 6 — BP splits, sampling, inputs, and calibration.
  • Supplement §§D.1–D.3 | pp. 16–17, Tables S25–S28 — Fixed-checkpoint inference gating and across-run confidence intervals.
AuthorsOpenReview ↗

Response to Reviewer F4NC (Part 3/4: Q3)

Q3: Please define L_c (formula, augmentations, temperature), the QRAFT fuser versus cross-attention fusion, and Comparisons I to III in Table 3, and expand FGQA and the other acronyms at first use.

A5: The contrastive term is symmetric in-batch InfoNCE:

Lc = 1/2 [ℓ(z,z′) + ℓ(z′,z)] ℓ(z,z′) = −(1/B) Σᵢ log [ exp(cos(zᵢ,z′ᵢ)/τ) / Σⱼ exp(cos(zᵢ,z′ⱼ)/τ) ]

Here, zi,zi′z_i,z'_i are embeddings of two independently augmented views of the same record, BB is the batch size, and τ=0.07\tau=0.07. Augmentations comprise amplitude scaling, offset, additive jitter, and temporal shifts; other records in the batch supply negatives. The spectral term uses the L1L_1 distance between log-magnitude spectra.

QRAFT-FUSE uses local quality scores to select measured ECG and PPG evidence before shared attention. Unreliable measurements are replaced by mask tokens while retaining their positions. The source tokens are concatenated, and shared attention models temporal and cross-source dependencies. Quality controls evidence selection rather than attention-logit weighting. The cross-attention fusion baseline exchanges information between modality features without this quality-based evidence selection at fusion time.

R1-Table3 crosses pretraining policy with fusion rule under the same downstream protocol. Within each fusion column, both checkpoints use the same inference settings. Cross-attention receives full ECG and PPG, and QRAFT-FUSE uses FGQA to select patches within the fusion module.

R1-Table3: Effects of pretraining and fusion on MIMIC-III SBP MAE (mmHg).

PretrainingCross-attentionQRAFT-FUSE
Random mask3.78 ± 0.133.56 ± 0.12
QRAFT3.65 ± 0.123.36 ± 0.10

With the fusion rule fixed, QRAFT pretraining reduces SBP MAE by 3.4%–5.6%. With pretraining fixed, the QRAFT-FUSE configuration, which combines FGQA-based evidence selection with shared attention, reduces SBP MAE by 5.8%–7.9% relative to the cross-attention baseline. These results indicate complementary contributions from reliability-conditioned pretraining and quality-aware fusion.

Comparison I retains coarse record-level quality screening before pretraining; II removes quality perception from target selection; III removes low-quality patches from both the evidence and reconstruction-target sets. In Eq. 1, eire_i^r denotes a patch embedding, tmaskrt_{\mathrm{mask}}^r a source-specific mask token, pirp_i^r a positional embedding, and gψrg_\psi^r the quality estimator. FGQA means fine-grained quality assessment; QA, quality assessment; AC Encoder, adaptive cross-encoder; MSFA, multi-scale feature aggregator; MHC, multi-head convolution; ECA, efficient channel attention; TSR-Loss, time-series reconstruction loss; PPG-PM, PPG precision masking; and PI, percentage improvement. Supplementary Section A.2 lists these definitions, and the simplified Figure 2 uses functional descriptions.

Revision locations.

  • Main text §§3.1–3.3 | pp. 4–5 — Notation, QRAFT-FUSE, reconstruction heads, and InfoNCE.
  • Supplement §A.2 | p. 3 — Implementation settings, Comparisons I–III, and acronym definitions.
  • Supplement §C.4 | p. 12 — Fusion inference settings in the factorial comparison.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank Reviewers F4NC, YTEV, and P1yQ, and the Area Chair, for their thoughtful and constructive feedback. We have revised the manuscript and supplement to clarify QRAFT's contribution, add targeted experimental evidence, and explain the evaluation protocols. Below, we summarize the core contributions of our work and the principal updates made during rebuttal; detailed responses and revision locations are provided in the individual reviewer replies.

Core Contributions of Our Work

  1. Reliability-conditioned masked modeling: QRAFT uses fine-grained signal reliability to select visible evidence and reconstruction targets, allowing reliable temporal and cross-source context to support prediction of withheld ECG and PPG patches.
  2. Controlled evaluation of reliability-aware selection: Difficulty-matched masking and a pretraining-by-fusion factorial evaluate the contribution of reliability-conditioned pretraining separately from generic difficulty selection and the fusion rule.
  3. Physiological evaluation across tasks and cohorts: We evaluate waveform reconstruction, blood-pressure transfer across three datasets, synchronized ECG–PPG fusion, and quality-stratified performance.

Updates of Experimental Results During Rebuttal

  • Main text §4.4; Supplement §§E.1–E.2, Tables S33–S35: Added CPC/JEPA comparisons using the same pretraining data and matched budgets, and frozen-encoder MIT-BIH heartbeat classification. QRAFT reaches 4.20 ± 0.22 mmHg SBP MAE on the five-run BP comparison and 95.9 ± 0.5% Accuracy / 88.1 ± 0.8% Macro-F1 on MIT-BIH over ten paired seeds.
  • Main text §4.2; Supplement §§B, C.2, C.5: Added shared-mask paired tests (0.82 to 0.72 versus random masking, p = 0.0023; 0.77 to 0.72 versus DM-Rand, p = 0.021), selected-patch difficulty distributions (absolute SMD 0.09–0.13), and the threshold-to-aggregate-mask analysis.
  • Supplement §§B.1–B.2 and C: Added known-location synthetic artifact evaluation (AUPRC 0.80), real-signal expert-score agreement and FGQA correlations, paired clean-reference recovery (9% lower RMSE), and physiological-feature preservation.
  • Supplement §§D–E: Added fixed-checkpoint inference-gating results, BP run-level intervals and constant-prediction controls, and the VitalDB pretraining-size ablation.

Updates of In-Depth Discussions During Rebuttal

  • FGQA and reproducibility: Clarified source datasets, record-level expert labels, adjudication, multi-instance learning, subject separation, and the status of the annotated-data release.
  • BP and ECG–PPG protocols: Specified subject-level BP splits, PPG-only input for BP transfer, absence of subject-specific calibration, and the paired ECG lead II / PPG- or PLETH-based fusion setting. The Discussion describes timestamp alignment and the remaining sampling and channel-timing variation.
  • Method and presentation: Defined the loss terms and notation, explained Equation (9), simplified Figure 2, clarified the fusion comparisons, and retained auxiliary backbone-transfer tasks in the supplement. Ethics approval and participant consent are stated in the revised manuscript.

The revised manuscript and supplement, updated on November 18, 2026, contain the corresponding changes, with revised passages marked in blue. We thank the reviewers and Area Chair for their time and constructive suggestions, which helped us strengthen the evidence and improve the clarity of the work.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.