acceptodds
Under review as a conference paper at ICLR 2027

Codebook-Anchored Prediction Across Physiological Views

17 views

Abstract

Physiological sensors capture related but non-equivalent views of the same cardiovascular interval. Cross-view reconstruction is difficult because the inverse mapping is ambiguous and continuous latent spaces provide no stable correspondence across sensors. We introduce \method, a codebook-anchored predictor that learns a shared four-level residual-vector-quantized (RVQ) tokenizer and freezes its latent vocabulary during cross-view prediction. Modality-specific encoders and decoders retain view-specific signal structure, while nearest-codeword projection constrains predicted target latents before waveform decoding. We use a fixed subject-disjoint split of MIMIC-IV Waveform v0.1.0 with 138/20/40 train/validation/test subjects. Against the strongest evaluated external baseline, reduces subject-level PPG-to-ECG RMSE from to and increases R-peak F1 from to . Under matched codebook parameter capacity, four-level RVQ reduces tokenizer RMSE from to relative to flat VQ. Using the same predictor checkpoint, nearest-codeword projection improves pooled-window F1 from to . The same interface supports 25-to-250/500-Hz ECG reconstruction with R-peak F1 of . In conclusion, these results show that our two-stage architecture supports both cross-sensor and cross-rate physiological reconstruction through a shared, frozen RVQ latent space within the evaluated MIMIC-IV cohort.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Predictive Representation Learning, Discrete Representation Learning, Cross-Modal Learning, Residual Vector Quantization, Multimodal Time Series

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_ZGe4OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes PhysioAnchor for cross-modal reconstruction from PPG to ECG, as well as reconstruction from low sampling rate to high sampling rate ECG. The method adopts a two-stage training approach. Firstly, a shared four-layer residual vector quantization codebook is learned, and modality-specific encoders and decoders are retained. Subsequently, the representation interface is frozen, and the potential predictors for the two tasks are trained separately. The prediction results are projected onto codewords and then decoded into ECG waveforms. Its main contribution is to organize two types of reconstruction tasks using a shared, frozen discrete target interface, while retaining task-specific mappings. The experiments use the subject isolation partition of MIMIC-IV. Compared to the strongest external baseline UA-P2E, the subject-level RMSE of PPG→ECG from 0.526 to 0.493, and the R peak F1 from 0.704 to 0.739. The paper also presents the ECG reconstruction results from 25 to 250/500 Hz. The ablation experiments support the role of residual quantization and pre-codeword projection before decoding in the current system. However, the validation is still limited to a single cohort and a single training for each configuration. The additional benefits of the shared codebook compared to the independent codebook have not been independently verified, and the results are not equivalent to the clinical diagnostic effectiveness.

Critical strengths

In the assessment of subject isolation, consistent improvement across various indicators was achieved compared to the stronger baseline. The paper generated windows after dividing the subjects and aggregated the test results by subjects, reducing the risk of evaluation bias caused by overlapping training and testing and related windows. This is the main basis for my positive evaluation of its empirical contribution. The benefits are reflected in waveform errors and peak detection, rather than being supported solely by individual visualizations or a single indicator.

Critical weaknesses

The added value of the core design has not yet been confirmed by a key comparison. The paper regards "shared, frozen discrete target interfaces" as its main contribution, but it does not compare the shared codebook with the independent modal codebook. The authors also explicitly admit that the causal contribution of the shared mechanism has not been isolated. At the same time, the "Continuous" in Table 3 merely cancels the decoding pre-projection of the same checkpoint, and is not a complete control from the first stage for training in a continuous representation. Therefore, this experiment supports the role of projection in the current system, but cannot prove that the discrete target space is superior to a reasonable trained continuous target space. This directly affects my judgment of the contribution of the method: the complete system achieves performance improvement, but this does not mean that its core shared and discretization design has demonstrated a substantive advantage over simpler alternative solutions compared to it.

The key ablation of the freeze strategy lacks an update mechanism consistent with the training objective. Equation (6) only calculates the error between the predicted latent variables and the target encoder output of the stopping gradient. The target codebook and decoder do not appear in the calculation path of this loss. However, Table 3 and the corresponding text report the results of "unfreezing the last-level quantizer" and "simultaneously unfreezing the decoder and the last-level quantizer", but do not explain how these components adapt through any additional loss or update rules. According to the disclosed calculation graph, simply removing the freeze is not enough to allow these components to obtain the gradient of this loss, so I cannot determine whether this experiment truly isolates the effect of the freeze strategy. This weakens the evidence supporting the core design and its reproducibility. It may be a missing training detail rather than an implementation error, but until it is clarified, it cannot be regarded as a reliable basis for supporting the acceptance.

Questions for authors

In some of the adaptation experiments in Table 3, how are the codebook and decoder updated? The predicted loss in Equation (6) only compares the predicted latent variables with the target encoder output at the stop gradient, and the target end codebook and decoder do not participate in the calculation of this loss. Please clearly define the complete training objective and parameter update mechanism for the "Last quantizer" and "Decoder + last quantizer" configurations. Are additional reconstruction losses, EMA, or other update rules used? Did the frozen control group adopt the same loss and training process, only changing whether the parameters are updated? Can the benefits brought by the shared codebook and the discrete target space be independently verified? The paper has explicitly acknowledged that there is no comparison between the shared codebook and the independent modal codebook; the existing "Continuous" control group only cancels the decoding pre-projection at the same checkpoint, not a continuous representation model starting training from the first stage. Can a matching control be provided with the independent modal codebook, the complete training of the frozen continuous autoencoder, and report the final PPG→ECG performance under the same data partition, static input, and comparable training budget? These two validations will directly affect my judgment on the contribution of the core method: whether the benefits indeed come from the sharing and discretization design, rather than mainly from the pre-trained target representation and decoder. If not provided, please clearly indicate which contributions have direct evidence and which still belong to unverified design motivations.

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

Summary: The paper reconstructs physiological signals across views using a shared codebook, evaluated with subject-disjoint splits. However, the practical objective is unclear, the novelty is limited, the shared-codebook claim is not isolated, and the evidence comes from a single small cohort.

Critical strengths

The paper includes subject-disjoint evaluation and controlled comparisons of several components. However, I do not find these strengths sufficient to outweigh the concerns below.

Critical weaknesses

Weaknesses

  1. The problem framing does not clearly establish the practical objective beyond waveform reconstruction. Although two reconstruction tasks are specified, the paper does not demonstrate why the reconstructed signals are useful for downstream applications or how the proposed framing advances existing approaches.
  2. The method is insufficiently specified in several important places. In particular, the masked cross-rate operator, encoder–decoder architectures, and construction of the high-rate reference signals require more detail to make the approach reproducible and the experiments interpretable.
  3. The technical novelty appears limited to combining established components, including residual vector quantization, latent prediction, and nearest-codeword projection. The paper does not sufficiently establish what new methodological principle emerges from this combination.
  4. The comparisons do not isolate the central claim that a shared codebook provides a beneficial cross-view reference space.
  5. The experimental evidence is limited to one cohort of 198 subjects and one training run per configuration. Without external validation, repeated runs, or downstream evaluation, the robustness and practical significance of the reported improvements remain uncertain.
Questions for authors

Questions:

  1. What is the practical objective beyond waveform reconstruction, and why are the reconstructed signals useful downstream?
  2. What evidence shows that a shared codebook is better than independently learned modality-specific codebooks?
Reviewer_eQBxOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PHYSIOANCHOR for PPG-to-ECG translation and cross-rate ECG reconstruction. Stage 1 trains modality-specific encoders and decoders around a shared four-level residual vector quantizer. Stage 2 freezes this interface and trains task-specific latent predictors, followed by nearest-codeword projection before ECG decoding. Evaluation uses a subject-disjoint MIMIC-IV Waveform split with 138/20/40 training/validation/test subjects.

Critical strengths

None.

Critical weaknesses
  1. Novelty. PHYSIOANCHOR primarily combines established components: residual vector quantization, modality-specific autoencoders, frozen-target latent prediction, Transformer-based regression, and nearest-codeword projection. The paper does not clearly establish a new quantization mechanism, prediction objective, or learning principle beyond this combination.

2.Limited empirical scope. All waveform experiments use a single MIMIC-IV Waveform v0.1.0 cohort containing 198 subjects, with only 40 held-out test subjects. Evaluation within this same cohort does not establish generalization across datasets, acquisition devices, or patient populations. Although the subject-disjoint split is appropriate, it tests generalization within the source cohort and provides no evidence of robustness to external domain shifts. It therefore remains unclear whether the gains from the learned RVQ vocabulary persist beyond this particular setting.

Questions for authors

What specific methodological advance distinguishes PHYSIOANCHOR from a combination of established RVQ autoencoding, frozen latent prediction, and nearest-codeword projection? Please identify the new mechanism or learning principle and explain what capability it enables beyond existing two-stage latent translation approaches.

Why is evaluation restricted to a single waveform dataset? Please evaluate the method and the strongest baseline on at least one independent paired ECG–PPG dataset with matched input and conditioning information.

Please provide controlled comparisons against independently learned modality-specific RVQ codebooks and a two-stage continuous-autoencoder predictor trained without quantization. These should evaluate downstream PPG-to-ECG reconstruction, rather than only Stage-1 autoencoding, to establish the contribution of the shared discrete interface.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.