acceptodds
Desk rejected at ICLR 2027

RAMP: Reliability Adaptive Masking for Pretraining Noisy Multimodal Physiological Signals

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

Abstract

Physiological signals from wearable and clinical monitors are often corrupted by motion artifacts, baseline drift, and contact noise, yet most self-supervised pretraining methods apply fixed masking policies regardless of local reliability. This can align reconstruction with artifact structure rather than physiological semantics. We present RAMP, a reliability-adaptive masking framework for noisy multimodal physiological pretraining. RAMP profiles each ECG and PPG patch using quality, corruption type, and spectral degradation, then uses a compact Reliability Policy Network to jointly select masking threshold, temporal granularity, and cross-modal fusion weights through differentiable relaxations. It reconstructs degraded regions from cleaner temporal and cross-modal context. On MIMIC-III and VitalDB, RAMP reduces pretraining RMSE by 16% over the strongest fixed-policy baseline, with the largest gains under asymmetric modality corruption. The pretrained encoder further achieves 93.9/93.9/93.2% mean ACC/macro-F1/ROC-AUC across six downstream physiological benchmarks spanning ECG, PPG, EEG, and accelerometry, outperforming reproduced baselines under a unified evaluation protocol. These results suggest that reliability-aware masking improves physiological representation learning under controlled synthetic corruption.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Self-supervised learning, Physiological time series, Multimodal representation learning, Adaptive masking, Signal quality, Robust pretraining, Wearable sensing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_KtTcOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAMP, a reliability-conditioned masked-pretraining framework for paired ECG and PPG signals. RAMP extracts patch quality, corruption-type, and spectral-degradation evidence, and uses a small Reliability Policy Network to select a masking threshold, temporal granularity, and cross-modal fusion weights; Reliability-Weighted Reconstruction upweights low-quality masked targets. The paper evaluates reconstruction on MIMIC-III and VitalDB under controlled corruption, then reports transfer to six downstream ECG, PPG, EEG, and accelerometry benchmarks.

Critical strengths
  • Well-Motivated Problem: The paper addresses a practical and well-motivated problem: heterogeneous noise in physiological signals, such as electrode-contact noise in ECG and motion-induced optical interference in PPG. Existing fixed masking strategies are ill-suited to handling local variations in signal quality, making the proposed approach a meaningful step toward more robust physiological signal representation learning.

  • Clear and Systematic Framework: The proposed framework is well organized, with clearly delineated modules for performance analysis, strategy generation, strategy-conditioned reconstruction, and downstream encoder transfer. This modular design makes the role of each component easy to understand and facilitates systematic evaluation of the overall approach.

  • Comprehensive Evaluation and Generalization: The paper provides extensive ablation studies on MIMIC-III and cross-dataset validation on VitalDB. Furthermore, its evaluation extends to cross-modal downstream tasks involving EEG and motion accelerometer signals, providing encouraging evidence of the generalizability and transferability of the learned representations.

Critical weaknesses

1. Dependence on Manual Annotations and Limitations in Generalization

The proposed framework relies substantially on manually annotated calibration data. Specifically, Appendix C.4 states that the corruption-type classifier requires 500 manually labeled calibration segments per modality, annotated by biomedical engineers. This introduces additional annotation costs and domain-specific prior knowledge, potentially limiting the framework's applicability across different datasets and clinical environments. More importantly, this dependency raises concerns about the framework's ability to achieve robust adaptation beyond predefined corruption-handling strategies. When deployed in previously unseen clinical settings or with novel wearable devices, corruption patterns and signal quality distributions may differ substantially from those observed during calibration. However, the paper does not investigate whether the proposed framework can generalize to such environments without additional manual annotations.

2. Coarse-Grained Cross-Modal Alignment and Insufficient Consideration of Physiological Differences

According to Appendix D, RAMP enforces a shared temporal granularity gg for ECG and PPG within each window. However, these modalities reflect fundamentally different physiological processes: ECG captures cardiac electrical activity, whereas PPG measures peripheral blood volume changes. Their temporal dynamics and morphological characteristics differ, and the pulse transit time (PTT) introduces a physiologically meaningful delay between cardiac electrical events and peripheral pulse responses. Applying a shared temporal granularity to both modalities may fail to account for these modality-specific characteristics and could potentially obscure fine-grained temporal relationships between them. The paper does not sufficiently justify why a shared granularity is appropriate for signals with different physiological origins, nor does it establish whether this design preserves meaningful cross-modal temporal dependencies.

3. Insufficient Evaluation of Robustness to Missing Modalities

The proposed RAMP framework appears to assume the simultaneous availability of ECG and PPG. In particular, the cross-modal weighting formulation, αECG+αPPG=1\alpha^{\mathrm{ECG}} + \alpha^{\mathrm{PPG}} = 1 , appears to presuppose that both modalities are available during training and inference. However, modality loss is common in real-world clinical monitoring and long-term wearable sensing, for example, when ECG electrodes lose contact or PPG sensors become displaced. In such scenarios, one modality may become entirely unavailable, intermittently corrupted, or temporally misaligned with the other. These conditions may undermine the assumptions underlying the cross-modal weighting mechanism and potentially lead to unstable predictions or degraded representations. The paper does not evaluate whether RAMP remains effective under complete modality absence, severe temporal asynchrony, or dynamically changing sensor availability, leaving its robustness in realistic deployment scenarios insufficiently established.

Questions for authors

Q1. How does the proposed framework perform when transferred to unseen datasets, clinical environments, or wearable devices without access to manually labeled calibration segments? Could the authors provide ablation studies comparing the full model with zero-shot transfer and annotation-free adaptation settings to quantify the extent to which expert annotations are necessary for robust generalization?

Q2. What is the physiological and empirical justification for enforcing a shared temporal granularity across ECG and PPG? Could the authors compare the current design with modality-specific temporal granularities or PTT-aware alignment strategies and report whether these alternatives improve representation quality and downstream performance?

Q3. How does RAMP handle cases in which ECG or PPG is completely missing, intermittently unavailable, or severely misaligned? Could the authors evaluate the framework under controlled modality-dropout ratios, single-modality inference, and varying degrees of temporal asynchrony, comparing its performance with corresponding single-modality baselines? In particular, does the cross-modal weighting mechanism remain well-defined when one modality is unavailable, and what modifications, if any, are required to ensure stable inference?

Reviewer_J1EWOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

To prevent self-supervised pretraining on physiological signals from reconstructing sensor noise rather than underlying physiology, this paper introduces RAMP, a framework that profiles patch-level signal quality, corruption type, and spectral degradation to dynamically adapt the masking threshold, patch granularity, cross-modal fusion weights, and reconstruction loss weights during pretraining.

Critical strengths

Interesting, unique problem that has been identified with asymmetric degradation across different modalities. However, this is somewhat under-explained throughout the paper, and only really explained in the first two paragraphs of the intro. Nonetheless, it is fine because ICLR is an ML venue.

Critical weaknesses

I have many concerns with this paper, with much of the paper contradicting itself. I cannot in good faith review this paper further as there are many systematic issues with the paper that call into question the reliability of the authors themselves. 

1. Contradictions on input modalities

  • On line 140, the authors state: "Let S={ECG,PPG}", and on Appendix line 1173, they state that EEG and Accelerometry are "modalities never seen during pretraining".

  • However, Figure 1 (line 054) and Figure 2 (lines 165–174, "3. Policy-Conditioned Pretrain") explicitly depict 4 modalities (ECG, PPG, EEG, ACC) with 4 parallel encoders (EECG,EPPG,EEEG,EACC) and 4 decoders (DECG,DPPG,DEEG,DACC) during pretraining.

2. Contradictions in how the Gumbel-sigmoid relaxation is used

  • On lines 203–205, the authors define a Gumbel-sigmoid relaxation with a_i in Eq. 3 and claim straight-through gradients are used to train the policy.

  • On line 215, the authors define the final fused representation, and on lines 230, 239, and 244 (as well as Algorithm 1 on Appendix lines 046–053), the forward pass and losses are defined using hard subset indexing (Mt{i:qim<}), completely omitting the Gumbel-sigmoid relaxation a_i hat from the computational graph.

3. Contradictions on whether FGQA/RPN are used during inference

  • On lines 222, the authors state: "After pretraining, the profiler and policy are discarded and downstream transfer uses the encoder alone, so adaptation changes the learned representation without introducing test-time policy inference."

  • On Appendix lines 1312, the authors state: "The FGQA quality estimator and RPN policy introduce inference overhead relative to plain MAE, which warrants consideration in latency-sensitive clinical deployment", and Table A10 (Appendix line 756) and line 799 explicitly report a 1.03x (+3%) inference time overhead and 1.02x memory overhead from computing the RPN policy at inference time.

4. Contradictions on sections and equations available in the paper

  • On Appendix line 394, the authors state: "OOD evaluation in Appendix C.10" … There is no section C.10 (Appendix C ends at C.8).

  • On line 299, the authors state: "corruption-injection protocol are in Appendix B" … Appendix B only covers dataset descriptions (B.1) and hyperparameter tables (B.2), omitting the synthetic corruption-injection protocol entirely.

  • On Appendix line 164, the authors state: "main paper Eq. 7" … There is no Equation 7 in the main paper (which ends at Eq. 6).

Questions for authors

Why are there so many discrepancies in the paper?

Reviewer_icrpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes RAMP, a reliability-adaptive framework for masked pretraining of physiological signals. It uses signal quality, corruption type, and spectral information to jointly adjust masking thresholds, temporal granularity, and cross-modal fusion weights, together with reliability-weighted reconstruction. Evaluation covers reconstruction on MIMIC-III and VitalDB and representation transfer across six downstream benchmarks.

Critical strengths
  1. Conditioning masking, temporal scale, and modality fusion on signal reliability provides a coherent and useful approach to pretraining with heterogeneous corruption.

  2. The controlled ablations help distinguish the contributions of reliability evidence, joint policy adaptation, and reconstruction weighting. Additional augmentation and robustness experiments strengthen the empirical evaluation.

Critical weaknesses
  1. The central motivation is not directly validated. The paper argues that fixed masking may encourage learning artifact structure rather than physiological semantics, but the experiments primarily demonstrate performance improvements rather than this failure mechanism. The connection between the motivation and the observed gains therefore remains insufficiently established.

  2. Reconstruction comparisons remain concentrated on masking-policy variants and MAE augmentation controls. These comparisons are useful, but provide limited evidence of an advantage over other reconstruction-based pretraining methods. Some comparison protocols also require clarification, particularly the differing MAE downstream results between Table 4 and Appendix Table A25.

Questions for authors
  1. What evidence directly supports the claimed tendency of fixed masking to learn artifact structure, particularly given that reconstruction targets are the signals before synthetic corruption?

  2. Are reconstruction comparisons with other relevant pretraining methods available? Please also clarify the protocol differences underlying the MAE results in Table 4 and Table A25.

Clarifications addressing these concerns could change my assessment.

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

Related papers

Discussion (0)

Sign in to comment.