acceptodds
Withdrawn from ICLR 2027

Pulse2View: Same-Sensor Cross-View Pre-Training for Patient-Level Multi-Disease Classification from Wearable PPG

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

Photoplethysmography (PPG) is a widely available biosignal in consumer wearables, enabling scalable, low-burden cardiovascular screening. However, labeled PPG datasets are scarce, and single-view self-supervised pre-training can underexploit the complementary morphology, rhythm, and time-frequency structure of PPG. To address this, we propose Pulse2View (P2V), a same-sensor cross-view pre-training framework aligning two views of each 30-second recording without an additional sensing modality. For each of three consecutive 10-second segments, the continuous wavelet transform (CWT) branch constructs a three-channel Ricker-CWT image from the exported PPG and its first- and second-order differences, while a Conv1D-BiGRU encodes the full sequence as a temporal (time-domain) view. Symmetric InfoNCE aligns the two views without disease labels. Both encoders are transferred and fused end-to-end downstream, while inference remains PPG-only. Pre-trained on 99,701 unlabeled 30-second PPG windows, P2V was evaluated on the private hospital-derived Clinical Wearable PPG Cohort (CWPC; 14,811 patients and 100,657 windows). It achieved the highest three-seed mean on all four primary patient-level metrics among evaluated models, including a coronary heart disease (CHD) AUROC of 0.7859 ± 0.0044 and a Macro AUROC of 0.7150 ± 0.0012, exceeding a public image-pretrained ConvNeXtV2-Tiny initialization by 3.2 and 1.4 AUROC points and the strongest public baseline, MOMENT-small, by 2.6 and 0.8 AUROC points. At 25% and 50% label budgets, P2V improves over the image-pretrained initialization and a two-CWT-view control, while achieving the highest CHD and Macro AUPRC among these configurations. Paired bootstrap and ablations support the contributions of cross-view alignment and the pre-trained temporal branch. On the public Liu PPG arrhythmia six-class dataset, P2V also achieves a window-level Macro AUROC of 0.9697 ± 0.0011. Together, these results support label-efficient cross-view pre-training for PPG-only patient-level multi-label classification.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
photoplethysmography, wearable sensing, self-supervised learning, contrastive learning, representation learning, cross-view learning

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

Pulse2View (P2V) presents a self-supervised learning approach that constructs two complementary representations of the same PPG signal to capture important morphology features. In particular, the first representation is in the image space where the PPG is transformed using CWT and encoder using a ConvNeXt. The second representation is a time series view where the PPG is encoded using 1D CNN + BiGRU. P2V is pre-trained using InfoNCE loss over these representations across 99K windows. After pre-training, the model is fine-tuned on 71K windows for multi-label classifications of 7 diseases. The validation demonstrates superior performance compared to prior baselines on a private dataset.

Critical strengths
  • P2V's use of multiple representations is interesting because it still relies on a single PPG sensor, making it useful for deployments.
  • Label efficient learning and the six class task external evaluation are useful experiments that demonstrate the efficiency and some generalization of P2V.
Critical weaknesses

(Experimental weakness) Unfair comparisons to prior SSL approaches: The closest baseline to P2V is TF-C. However, only P2V is pre-trained on 99.7K unlabelled pre-training. For fair comparison, TF-C should also be trained on the same pre-training data to ensure that the performance gains are because of pre-training method and not the data. Unlike public models which can be evaluated directly, TF-C is an SSL approach, hence it should be evaluated in the way mentioned above.

(Experiment weakness) Single representation temporal SSL baseline: The claim that single sensor at inference time does not mean that pre-training requires a single sensor requires an SSL baseline that uses a single representation. This comparison will support that multiple representations in SSL are better than a single one. As PPG is a time series, using a single representation approach like SimCLR for PPG is necessary. This baseline is more important that pre-trained baselines because P2V is presented as an SSL approach.

(Experimental weakness) Pre-training vs. Fine-tuning distributions: The paper states that the pre-training data is split prior to constructing CWPC. This implies that they are part of the same study and could have the same devices and protocols. Consequently, P2V pre-training and evaluation distributions could be similar. If so, P2V is in an easier in-domain setup compared out-of-distribution for other baselines.

(Experimental weakness) Fine-tuning vs Probe: A major approach to evaluate the effectiveness of SSL approaches transfer to downstream performance is a linear probe (multi-label probe in P2V's case). This would indicate if the representations learned from multiple views is actually useful for downstream tasks. However, the experiment results rely on fine-tuning, which update the pre-trained model. Therefore, it is unclear whether downstream performance is driven by fine-tuning or P2V's multi-representation pre-training.

(Results) Confidence Intervals: Providing bootstrapped CIs from a single seed would be more useful than 3-seed averaging in Table 2. These CIs will help the reader understand if P2V's improvements are meaningful given the smaller sample size of CHD labels.

Questions for authors

These questions and comments did not affect my score:

(1) What are the inputs to the traditional controls (e.g., RF, LR)? Is feature engineering performed prior to training these ML models.

(2) The positive-class weight wc∈[1,20]w_c \in [1,20] , wouldn't these value dramatically increase the loss? Wouldn't it better scale it between 0 to 1 to obtain ratio?

(3) The Experimental Setup can be better presented to clarify which models are trained with what data. SSL approaches can be separated from PPG or Time series native models.

(4) Why not evaluate on other large PPG datasets such as MIMIC, VitalDB, or others? P2V is not restricted to multi-label classification as that is only downstream.

Reviewer_isAnOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper developed Pulse2View for patient-level disease prediction from wearable PPG alone. It constructs two views: CWT images from three 10-second segments, encoded by a shared ConvNeXtV2, and the full waveform, encoded by a Conv1D-BiGRU. The encoders are pretrained with symmetric InfoNCE and then jointly fine-tuned after feature concatenation.

Critical strengths

The idea of coupling local CWT time-frequency features with the full PPG waveform is well motivated. The frozen-feature results in the supplemental materials also suggest that cross-view pretraining improves the CWT representation itself.

The experimental evaluation is thorough. CWPC includes 14,811 patients, and the paper documents separation of pretraining and downstream patients, patient-disjoint splits, three seeds, and paired statistical analysis. P2V achieves the best mean on the four designated primary patient-level metrics among the evaluated models, with additional per-disease and limited-label results.

Critical weaknesses

Prior work such as TF-C already used time-frequency self-supervision and downstream fusion. P2V differs mainly in its CWT view construction and encoder choices, so the methodological novelty is limited.

While the ablations are useful, there is no comparison that keeps the same two-branch architecture and unlabeled data but removes cross-view alignment. It is thus difficult to single out the contribution of the InfoNCE objective itself.

The performance gains are not consistent. Compared with MOMENT-small, the Macro AUROC improvement is only 0.0083, while AUROC decreases for arrhythmia and stroke. At 25% and 50% label budgets, CHD AUROC also does not exceed CWT-only. TF-C, a particularly relevant baseline, performs unexpectedly poorly and its training protocol needs clarification.

Finally, the seven-disease results come from one private cohort; the external experiment addresses a different arrhythmia task and does not establish cross-cohort generalization for the main task.

Questions for authors

With the same CWT + Conv1D-BiGRU architecture and unlabeled data, how does performance change without cross-view InfoNCE (or with the branches pretrained separately before fusion)? Can the current results distinguish the benefit of the two-branch architecture from that of cross-view alignment?

Which pretrained weights or data, fine-tuning procedure, and classification thresholds were used for TF-C? How do the authors explain its Macro AUROC of 0.5980 and Macro F1 of 0.016, both below logistic regression?

For the public six-class arrhythmia experiment, please clarify the source of the end-to-end training data, the train/validation/test split, and how the 10-second recordings were adapted to the proposed 30-second, three-segment CWT pipeline.

Are results for public pretrained baselines such as MOMENT-small available at the same 25% and 50% label budgets? They would help assess the claimed label efficiency.

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

The authors propose Pulse2View (P2V), a self-supervised pre-training framework for PPG that aligns two views of the same 30-second recording without using any additional sensing modality. The first view is a time–frequency image: each of three consecutive 10-second segments is converted into a three-channel Ricker-CWT image built from the PPG and its first- and second-order differences, then processed by a vision encoder (ConvNeXtV2-Tiny). The second view is the full waveform, encoded with a Conv1D-BiGRU. The two views are aligned with a symmetric InfoNCE objective on ~100k unlabeled windows, and both pre-trained encoders are then transferred and fused for end-to-end fine-tuning on downstream tasks. The main evaluation is a private, patient-level seven-disease benchmark (CWPC, 14,811 patients) with CHD AUROC as the primary metric. On it, P2V reports the best mean on all four headline metrics against seven public time-series/PPG foundation models and three classical baselines.

Critical strengths
  1. The motivation to use heterogeneous views of the same PPG to train a foundation model makes sense, and the paper offers direct evidence for it.
  2. The evaluation protocol is careful. Splits are patient-disjoint, patients in the pre-training corpus are excluded from the downstream cohort, predictions are aggregated and scored at the patient level, and AUPRC is reported alongside AUROC.
Critical weaknesses
  1. I find the benchmarking to be rather thin. The headline claim rests on a single private benchmark, and it is not a particularly discriminative one. A basic logistic regression already reaches 0.721 CHD AUROC and 0.674 Macro AUROC (vs. 0.786 and 0.715 for P2V), and it outperforms five of the seven foundation-model baselines on Macro AUROC. Some of this may come down to how the baselines were set up. The arrhythmia benchmark helps, but its comparisons are included only in the supplemental material and the main text reports P2V alone. For a PPG foundation model paper, using more standard benchmarks would strengthen the paper significantly. Consider tasks such as heart rate estimation, blood pressure estimation, etc. Both PaPaGei and Pulse-PPG, which the authors used as baselines, have evaluated their models on those tasks. Relatedly, all downstream results use end-to-end fine-tuning with a task-specific head. This is a perfectly sound choice, but I'd also add linear probing results for the full model and for the baselines, since linear probing is very standard across the PPG FM literature (both PaPaGei and Pulse-PPG evaluate their models with a linear probe on frozen encoders). The frozen readout in Table S7 covers only the CWT encoder and the internal variants.
  2. I'm not particularly convinced by the novelty framing. Contribution (1) argues for heterogeneous representation, but this premise is already well established. TF-C (Zhang et al., 2022), which the authors cite, jointly learns time- and frequency-domain representations of the same series. Contrastive alignment between a raw-waveform encoder and a time–frequency encoder of the same signal has also been explored for audio (Wang & van den Oord, 2021) and EEG (Kumar et al., 2022). More importantly, the CWT branch appears to be essentially the tokenization used in NormWear (Luo et al., 2025). NormWear computes the first- and second-order derivatives of the signal, applies a CWT with scales 1–64 to the raw signal and both derivatives, and stacks the resulting scalograms into a three-channel image. P2V's CWT view follows the same recipe, down to the 64 scales (Supp. Sec. 9). NormWear is also one of the paper's own baselines (Table 2), but it is cited only as a multi-signal foundation model, not as the source of this construction.
  3. Label-efficient learning would tell a much better story if it were compared to other baselines instead of just internal ablations. It would be very interesting if, for example, P2V learns more efficiently than other FMs like Pulse-PPG or MOMENT-small, either compute-wise or data-wise or both. An internal comparison only shows that the components work but does not deliver a strong claim on efficiency. Related to this, the pre-training corpus (99,701 30-second windows, roughly 830 hours) is small compared to many other PPG foundation models. PaPaGei, for instance, was pre-trained on over 57,000 hours. Could there be something to claim about pre-training data efficiency (i.e., delivering good quality representations using only a very small pre-training budget)? Assuming that the authors have saved the intermediate checkpoints, it'd be interesting to understand how the downstream performance progressed over the course of pre-training. And I think that it would be a stronger paper if the authors can establish that 1) P2V performs well on a wider range of benchmarks/tasks and 2) requires much fewer pre-training data than other FM counterparts in doing so.

Citations:

  1. Luo, Y., Chen, Y., Salekin, A., & Rahman, T. (2025). Toward foundation model for multivariate wearable sensing of physiological signals. arXiv preprint arXiv:2412.09758
  2. Zhang, X., Zhao, Z., Tsiligkaridis, T., & Zitnik, M. (2022). Self-supervised contrastive pre-training for time series via time-frequency consistency. arXiv preprint arXiv:2206.08496
  3. Wang, L., & van den Oord, A. (2021). Multi-format contrastive learning of audio representations. arXiv preprint arXiv:2103.06508
  4. Kumar, V., Reddy, L., Sharma, S. K., Dadi, K., Yarra, C., Raju, B. S., & Rajendran, S. (2022). mulEEG: A multi-view representation learning on EEG signals. arXiv preprint arXiv:2204.03272
Questions for authors

For the main questions, please refer to the weakness section above. Here are some additional (less important)

Clarifications

  1. In line 076 the authors used the phrase "PPG-only deployment". Are the authors specifically discussing deployment (i.e., test-time inference) or training in this passage? I'm asking because PPG-only deployment is the default for PPG foundation models, including ones that pre-train with other modalities such as ECG.

Suggestions

  1. It might read slightly better to put the supplemental PDF as the appendix.
  2. The per-disease results (Tables S3–S4) are worth moving into the main text, since they show where the gains actually come from.

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

Related papers

Discussion (0)

Sign in to comment.