- 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.