acceptodds
Under review as a conference paper at ICLR 2027

Unlocking the Specialization Capability of PPG Foundation Models for Wearable Signals through Structured Pruning and Adaptive Fine-Tuning

Abstract

Photoplethysmography (PPG) plays a central role in wearable health monitoring and clinical decision support. However, existing approaches to universal PPG representation learning are predominantly trained on high-quality public benchmark datasets and often struggle to adapt to sparse and low-quality real-world datasets, thereby limiting their generalization to complex clinical tasks and heterogeneous patient populations. Surprisingly, even after full fine-tuning, PPG foundation models cannot consistently outperform general-purpose time-series models. Therefore, a key question is how to effectively adapt PPG foundation models to target prediction tasks. Through empirical studies on various PPG foundation models, we find that pretrained models often exhibit inherent sparsity and computational redundancy, suggesting that PPG foundation models have learned to activate task-relevant network substructures for different prediction tasks. To preserve this valuable prior knowledge, we propose a structured pruning-and-fine-tuning framework, **SPAFT**, which regularizes the adaptive fine-tuning process by constraining it to a more relevant and compact parameter space based on parameter importance and structural redundancy. Extensive experiments across four PPG foundation models, five benchmark datasets, and ten downstream tasks demonstrate that SPAFT achieves win rates of 82.8% and 87.5% on regression and classification tasks, respectively. This “prune-then-finetune” paradigm often enables PPG foundation models to achieve state-of-the-art performance, laying a promising foundation for the real-world clinical deployment of PPG foundation models and the development of personalized PPG foundation models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Photoplethysmography, PPG Foundation Model, Structured Pruning, Adaptive Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_JRyBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SPAFT, a method to prune and then fine-tune PPG foundation models. Initially, the paper identifies that existing CNN based PPG FMs have significant redundancies and many blocks and filters do not contribute toward task prediction. Therefore, SPAFT first identifies redundancies in resnet blocks by structurally comparing cosine similarities of the filters within a block. Then, these are pruned and fine-tuned. Experiments demonstrate that SPAFT outperforms baselines across both classification and regression with rates of 82.8% and 87.5%, respectively.

Critical strengths
  • SPAFT is well-motivated through experiments with PPG FMs. It addresses an important problem of adapting FMs for different health tasks that could require different PPG characteristics for prediction.
  • Comprehensive baselines: The paper compares SPAFT against well-justified baselines across four PPG foundation models and many fine-tuning paradigms, demonstrating an extensive experimental design.
Critical weaknesses

Addressing these weaknesses will help me reconsider my score:

  • Statistical Significance: The results reported in Table 1 and Table 2 are single values which are close. For example, for Ring-PPG HR, SPAFT obtains 58.447 for AnyPPG and the 2nd best is 59.285. Consequently, it is unclear if SPAFT's improvement is meaningful. In particular, the win rates are calculated based on which method achieves the best numerical result. This approach treats negligible differences as wins, which is not meaningful. To address this issue, include statistical tests between the top 2 results to identify an useful win rate metric.
  • Dataset provenance: Dataset details such as the total number of samples, people, devices are necessary to contextualize the result. This will help the reader understand whether the improvements are meaningful. For example, an improve of 0.1 on 10 people is less likely to be significant compared to one evaluated across 1000. Furthermore, the paper emphasizes heterogeneous deployments, however, four of the evaluation tasks are on a private dataset and most tasks focus on either heart rate or blood pressure. I recommend either tempering the claims on the cardiovascular domain or diversifying the evaluation by adding more tasks. For example, glucose or respiration rate from the already used BUT-PPG dataset, or depression from the ECSMP dataset (https://www.sciencedirect.com/science/article/pii/S2352340921009355).
  • Overclaims:
    • (1) Efficiency: The experiments in the paper do not support the efficiency claims. Perhaps, these results are in the appendix, which is missing. Also, I recommend including the primary supporting evidence in the main text instead of the appendix.
    • (2) Task-relevant substructures: The paper indicates that SPAFT finds task-relevant substructures. While the prune-then-finetune approach could learn these structures, additional evidence such as representations before and after SPAFT that demonstrate the substructure is necessary to substantiate this claim (e.g., a heart rate substructure).
    • (3) Case Study: In Figure 4, QLoRA does show inter-class separation, hence saying it fails to distinguish the two waveforms is an overclaim. Visually, SPAFT is only marginally better. Could you provide quantitative metrics to support this case study?
Questions for authors

Questions and suggestions that will affect my assessment are in the Weaknesses section.

3 comments
AuthorsOpenReview ↗

Thank you again for your time and review

Dear Reviewer JRyB,

Thank you again for your time and review. Have we addressed your concerns? We would appreciate the opportunity to further clarify any remaining perceived issues as you finalize your assessment.

Best,

Authors of submission 25063

Thank you for the additional experiments and clarifications. My concerns have been addressed and I have increased my score.

We sincerely thank the reviewer for carefully reviewing our work and increasing the score. We appreciate the constructive feedback, which helped us strengthen both the analysis and the experiments.

Reviewer_ThcaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SPAFT, a ‘Pruning + Adaptive Fine-tuning’ framework designed to improve the downstream task adaptation and computational efficiency of PPG FMs. SPAFT consists of two main stages. First, task-structure collaborative pruning combines task-aware importance estimation with graph-based structural redundancy analysis. Second, importance-guided adaptive fine-tuning re-estimates the importance of the remaining filters after pruning and allocates a limited fine-tuning budget across blocks according to the concentration of highly important filters. Evaluations on 4 PPG FMs across multiple downstream tasks and datasets show the efficiency of SPAFT.

Critical strengths

Overall, the paper is well-organized and thoroughly evaluated, with most claims clearly articulated and supported by experimental evidence.

  1. Well-motivated task-structure collaborative pruning strategy. The proposed approach effectively integrates task importance and structural redundancy to guide filter pruning across different layers.

  2. Adaptive fine-tuning after pruning. SPAFT re-estimates filter importance after pruning and adaptively allocates fine-tuning budgets to task-relevant components, improving parameter efficiency.

  3. Comprehensive experimental evaluation. The experiments cover multiple PPG foundation models, datasets, and downstream tasks, with comparisons against 14 baselines spanning five categories, including conventional adaptation, parameter-efficient fine-tuning, structured pruning, and foundation models.

Critical weaknesses

Major weaknesses

  1. Limited PPG-specific methodological novelty. Although SPAFT is motivated by the characteristics of PPG foundation models (lines 53-60), its core components are largely domain-agnostic. The proposed framework appears to be a general structured pruning-and-fine-tuning approach rather than a method specifically tailored to PPG signals. The authors should better clarify how the unique characteristics of PPG data motivate the proposed design choices and what distinguishes SPAFT from general-purpose pruning and fine-tuning frameworks.

  2. Issues with the adaptive fine-tuning budget allocation (Equations 7–8).

(1) Inconsistent cross-layer importance comparison. The authors avoid directly comparing filter importance scores across layers during pruning, but Equation (7) relies on a global Top-M ranking of these scores for fine-tuning budget allocation. Layer-width normalization does not resolve potential differences in importance scales across layers. This inconsistency requires further justification.

(2) Mathematical inconsistency in Equation (8). Since the allocation weights satisfy ∑ℓωℓ=1\sum_\ell \omega_\ell=1, the proposed formula qℓ=1+Mωℓq_\ell=1+M\omega_\ell yields ∑ℓqℓ=M+L\sum_\ell q_\ell=M+L, contradicting the stated total budget (M). The formula should be clarified or corrected, potentially to qℓ=1+(M−L)ωℓq_\ell=1+(M-L)\omega_\ell, with appropriate integer rounding and capacity constraints.

  1. Missing No-Pruning Ablation. The paper lacks a comparison with selective fine-tuning on the original, unpruned model, making it unclear whether the performance gains come from pruning or selective fine-tuning. This weakens the justification for the proposed prune-then-finetune framework.

  2. Sensitivity to the similarity threshold. The cosine similarity threshold τ\tau directly affects the graph structure and redundancy estimation, potentially influencing pruning decisions. The authors should provide a sensitivity analysis to justify the choice of τ\tau and demonstrate the robustness of SPAFT.

Minor weaknesses

Figure 1 d: Low-Conttibution-> Low-Contribution

Most of my concerns have been addressed. However, I believe the paper would benefit from a more detailed discussion of its technical differences from existing task-aware structured pruning methods. In my view, as also noted by Reviewer NdtY, the proposed approach appears to be more of an adaptation and application of existing techniques.

Questions for authors

Please primarily refer to Question 1 in Section Critical Weaknesses. Given that SPAFT appears to be a general pruning-and-fine-tuning framework, could the authors clarify how the proposed approach differs from existing pruning and fine-tuning methods? I would be willing to raise my score if the authors can adequately address these concerns.

4 comments
AuthorsOpenReview ↗

Thank you again for your time and review

Dear Reviewer Thca,

Thank you again for your time and review. Have we addressed your concerns? We would appreciate the opportunity to further clarify any remaining perceived issues as you finalize your assessment.

Best,

Authors of submission 25063

AuthorsOpenReview ↗

Further follow-up before the discussion concludes

Dear Reviewer Thca,

Thank you very much for your careful and thorough review of our paper. As the discussion phase will conclude on November 18, we would like to briefly summarize the major revisions made in response to your comments:

  • PPG Data-Inspired Design Choices: We specifically design the pruning and fine-tuning stages based on the characteristics of downstream PPG data (Section 4).
  • Fundamental Differences of SPAFT: We clarify the fundamental differences between SPAFT and conventional pruning-and-fine-tuning frameworks in terms of Task-Structure Collaborative Pruning, Importance-Guided Selective Filter-Path Fine-Tuning, and Layer-Width-Normalized Adaptive Budget Allocation (Response [2/6]).
  • Inconsistent Cross-Layer Importance Comparison: We validate the effectiveness of our design from two perspectives: Different objectives require different filter selection strategies and additional experiments (Section 4.2 and Table 4).
  • Mathematical Inconsistency in Equation 8: We have corrected the equation.
  • Missing No-Pruning Ablation: We have added the no-pruning ablation experiment (Table 4).
  • Sensitivity to the Similarity Threshold: Please refer to Appendix C.6 in the original supplementary material.

If these revisions have addressed your concerns, we sincerely hope you will take them into account when preparing your final assessment and consider updating your evaluation accordingly.

Best regards,

The Authors

AuthorsOpenReview ↗

Any open questions before the discussion ends today?

Dear Reviewer Thca,

Thank you again for your time, and we sincerely apologize for reaching out once more during such a busy period. As the discussion period closes today, we would like to politely follow up one last time. We sincerely hope that our responses and revisions have addressed your concerns; if so, we would greatly appreciate it if you could take them into consideration in your final rating. We would be more than happy to provide further clarification if needed.

Best regards,

The Authors

Most of my concerns are addressed. I will raise the rating.

Reviewer_NdtYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SPAFT, a pruning and finetuning framework for adapting PPG foundation models to downstream tasks. It adapts and combines existing pruning methods, most notably Less-is-More and SRR-GR. It then re-estimates importance to allocate a limited fine-tuning budget across the pruned network. The authors evaluate four pretrained PPG FMs on five datasets for regression/classification tasks, reporting improvements over conventional finetuning and adaptation.

Critical strengths

As far as I know, this paper is the first application of pruning and finetuning for PPG foundation model. Since PPG data have high redundancy, the application of pruning is intuitive. The experiments also seem to be adequately sufficient.

Critical weaknesses

Since this paper is conceptually closely related to Less-is-More, which is positioned as a framework for general time series foundation models, I will this paper more as an adaptation and application. Hence, the novelty is moderate.

Questions for authors
  1. What are the Pearson correlation coefficients for the regression tasks, particularly SBP and DBP estimation, and provide MAE/MSE stratified by low, normal, and high blood pressure? In my opinion, these metrics can be more clinically relevant than overall MAE/MSE score, and PPG models can struggle in both.
  2. Please clarify how Less Is More is implemented as a baseline. Specifically, how is it adapted to the CNN-based PPG foundation models? What structures are pruned, how are importance scores computed and updated, and which parameters are subsequently fine-tuned?
4 comments
AuthorsOpenReview ↗

Thank you again for your time and review

Dear Reviewer NdtY,

Thank you again for your time and review. Have we addressed your concerns? We would appreciate the opportunity to further clarify any remaining perceived issues as you finalize your assessment.

Best,

Authors of submission 25063

AuthorsOpenReview ↗

Further follow-up before the discussion concludes

Dear Reviewer NdtY,

Thank you very much for your careful and thorough review of our paper. As the discussion phase will conclude on November 18, we would like to briefly summarize the key revisions made in response to your comments:

  • Differences from Less-is-More: We clarify the differences in terms of four aspects: target model architectures, Task-Structure Collaborative Node Pruning, Importance-Guided Selective Filter-Path Fine-Tuning, and Layer-Width-Normalized Adaptive Budget Allocation (Appendix A.3).
  • Pearson Correlation Coefficients for Regression Tasks: We evaluate PCCs for SBP and DBP regression on the UCI-BP and Ring-PPG datasets. Our method achieves the highest PCC in the vast majority of cases (Appendix C.2).
  • Results Across Different Blood Pressure Levels: We compare our method with four representative approaches on the UCI-BP dataset. Our method achieves the best prediction performance across different blood pressure groups (Tables 15 and 16).
  • Less-is-More Implementation Details: Please refer to Response [4/4].

If these additional clarifications and results have addressed your concerns, we would sincerely appreciate it if you could take these revisions into account and reflect them in your final assessment.

Best regards,

The Authors

Reviewer_NdtYOpenReview ↗

Response

I want to thank the authors for addressing my concerns and posting additional results. I have updated my score accordingly.

We sincerely thank the reviewer for carefully reviewing our work and increasing the score. We appreciate the constructive feedback, which helped us strengthen both the analysis and the experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.