acceptodds
Under review as a conference paper at ICLR 2027

PULSE: Physiology-Anchored Set Encoding for Variable-Lead ECG Representation Learning

Abstract

Most ECG representation models assume fixed lead configurations, limiting robustness to variations in lead availability, ordering, and signal degradation. In real-world deployment, ECG lead configurations vary widely across acquisition systems, from standard 12-lead recordings to single-lead wearables. We introduce PULSE, a physiology-anchored set encoding framework for variable-lead ECG representation learning, designed to learn from arbitrary subsets and orderings of the 12 standard ECG leads. The framework encodes each lead independently without explicit lead-identity embeddings and then combines the leads through two parallel branches: one models interactions between leads with a single pairwise cross-lead attention block and pools them by permutation-invariant set aggregation, while the other pools per-lead representations that stay unchanged by which other leads are present. It is pretrained with a self-supervised objective: given randomly selected lead subsets with temporal masking, it predicts fixed morphology and rhythm descriptors computed from the corresponding clean signals. A single fine-tuned model then supports varying lead configurations without configuration-specific retraining. Comprehensive experiments across multiple ECG datasets and lead configurations demonstrate that PULSE achieves competitive or superior performance to existing self-supervised ECG representation-learning methods, while remaining robust across variable-lead settings. With 0.43M parameters, PULSE runs on a Raspberry Pi 4B, processing a 10-second ECG in about 90ms more than faster than real time, with the same accuracy as on a workstation. These results show that variable-lead representation learning, compact modeling, and edge deployment can be achieved within a single ECG encoder.

Primary area:
learning on time series and dynamical systems
Keywords:
Electrocardiogram (ECG), Self-Supervised Learning, Variable-Lead Representation Learning, Permutation-Invariant Set Encoding, Physiological Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bCZdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PULSE, a compact self-supervised framework for variable-lead ECG representation learning. It treats available leads as an unordered set and combines a shared intra-lead encoder with two branches: a Cross-Lead Branch for inter-lead interaction and an Anonymous Branch for subset-stable per-lead encoding. Pretraining predicts fixed morphology and rhythm descriptors from temporally masked, randomly sampled lead subsets. The model is evaluated on PTB-XL, CPSC2018, and PhysioNet2017 under linear probing, fine-tuning, reduced/random lead availability, and lead reordering. It has 0.43M encoder parameters and is also deployed on a Raspberry Pi 4B.

Critical strengths
  1. PULSE addresses variable lead availability and ordering, which are common across clinical ECG systems, portable devices, and wearables.

  2. Shared per-lead encoding, cross-lead interaction, and symmetric aggregation provide a principled way to support variable-size inputs and invariance to input ordering.

  3. The linear-probe results are consistently strong relative to the selected self-supervised baselines, suggesting that the pretrained representation contains useful downstream information.

  4. The 0.43M-parameter encoder and measured Raspberry Pi latency, memory use, and numerical consistency provide credible evidence of edge-deployment feasibility.

Critical weaknesses
  1. The relationship between permutation invariance and lead identity needs stronger justification. Permutation invariance does not necessarily require removing lead identity: a set encoder can remain invariant to the ordering of lead–signal pairs while retaining a lead-ID feature. Since ECG leads correspond to distinct anatomical projections, it is unclear whether fully anonymous encoding is the most appropriate inductive bias.

  2. Anonymous encoding may limit the preservation of clinically relevant spatial information. Lead identity can be important for interpreting regional ST–T changes, infarct location, electrical axis, hypertrophy, and conduction abnormalities. A comparison with an otherwise identical permutation-invariant model using lead-ID or anatomical embeddings would be important for determining whether the proposed design improves robustness without sacrificing spatially relevant information.

  3. The central claim that one fine-tuned model handles arbitrary lead configurations without retraining is not yet sufficiently clear from the reported protocol. The paper should state explicitly whether the 12-, 6-, and 1-lead results use the same checkpoint, especially because the appendix distinguishes between “12-lead” and “variable-lead” models. If different configurations use separately fine-tuned models, the current evidence would not fully support this claim.

  4. The PTB-XL splitting protocol requires clarification. The paper appears to use random stratified splits rather than the official patient-wise folds. Because PTB-XL may contain multiple recordings from the same patient, record-level splitting could introduce patient overlap between training and test sets and complicate comparison with prior work. Results using the official or another patient-disjoint split would strengthen the evaluation.

  5. The construction of the single-label downstream tasks is insufficiently specified. PTB-XL and CPSC2018 may contain multi-label or multi-diagnosis records, whereas the paper states that every recording has one label. The filtering or label-reduction procedure should be described in detail. It would also be helpful to confirm that the quoted ST-MEM results use compatible labels and splits.

  6. The baseline coverage does not yet fully isolate the source of the reported gains. Particularly informative additions would include supervised training from scratch, a simple DeepSets or mean-pooling encoder, a lead-ID-aware permutation-invariant model, and a recent channel-agnostic or variable-lead ECG method. These comparisons would help distinguish the contributions of set encoding, random lead sampling, and physiology-anchored pretraining.

  7. More comprehensive ablations and statistical analyses would strengthen the conclusions. The current experiments do not fully examine cross-lead attention versus simpler pooling, the fusion/readout components, masking strategies, or the lead-sampling distribution. In addition, some fine-tuning differences are comparable to the reported standard deviations. Confidence intervals or paired significance tests would help determine whether these improvements are robust.

  8. The construction of pretraining targets under variable lead subsets should be clarified. It is unclear whether shared morphology, rhythm, and global descriptors are computed from only the retained leads or from the complete 12-lead recording. In the former case, the target may vary across subsets of the same recording; in the latter, the model predicts information derived partly from unavailable leads. Clarifying this choice and providing an ablation would improve the interpretation and reproducibility of the objective.

Questions for authors
  1. Lead identity and permutation invariance: Could the authors clarify why lead identity is removed rather than included as an element-level feature? Since symmetric set aggregation remains permutation-invariant when each signal is paired with its lead ID, could the authors compare PULSE with an otherwise identical model using lead-ID or anatomical/geometric lead embeddings?

  2. Clinically relevant spatial information: How does Anonymous encoding affect diagnostic tasks that depend on lead location, such as regional ST–T changes, myocardial infarction localization, electrical-axis abnormalities, hypertrophy, or conduction disturbances? If possible, could the authors provide per-class results or a targeted evaluation comparing anonymous and lead-aware representations?

  3. Single model across lead configurations: Do the reported 12-, 6-, 1-, and random-lead results all use the same fine-tuned checkpoint? Please clarify the lead-sampling procedure during pretraining, linear-probe training, fine-tuning, and testing, as well as the distinction between the “12-lead model” and “variable-lead model” in the appendix.

  4. PTB-XL data splitting: Does the PTB-XL evaluation use the official patient-wise folds or otherwise guarantee patient-level separation between training, validation, and test sets? If the current evaluation uses record-level random splits, could the authors provide results using the official or another patient-disjoint protocol?

  5. Single-label task construction: How are multi-label or multi-diagnosis recordings in PTB-XL and CPSC2018 converted into the reported single-label tasks? Please describe any filtering, label prioritization, or other reduction procedure, and clarify whether the quoted ST-MEM results use compatible label definitions and data splits.

  6. Baseline coverage: Could the authors provide comparisons with the most informative missing controls, particularly (i) supervised training from scratch, (ii) a simple DeepSets or mean-pooling set encoder, (iii) a lead-ID-aware permutation-invariant model, and (iv) at least one recent channel-agnostic or variable-lead ECG method?

  7. Ablations and statistical significance: Could the authors further isolate the contributions of cross-lead attention, pooling/fusion components, masking strategies, and the lead-sampling distribution? For improvements comparable to the reported standard deviations, could the authors provide confidence intervals or paired statistical tests?

  8. Pretraining targets under variable lead subsets: Are the shared morphology, rhythm, and global descriptors computed from only the retained leads or from the complete 12-lead recording? How does this choice affect target consistency across different subsets of the same recording, and could the authors provide an ablation comparing the alternatives?

Reviewer_9gW7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PULSE, a lightweight self-supervised representation learning framework for variable-lead ECG, to address the issues that existing ECG models rely on fixed lead configurations and struggle to adapt to missing leads and changes in lead order. The method treats ECG leads as an unordered set, employs a parameter shared single-lead encoder to extract temporal features, and uses two parallel branches to model cross-lead interactions and preserve stable single-lead representations, respectively. Symmetric aggregation is used to obtain a fixed-dimensional representation that is invariant to lead order. During pretraining phase, PULSE extracts morphological and rhythm statistical descriptors from clean ECGs and combines random lead sampling, temporal masking, and contrastive learning and regression losses to guide the model to learn physiologically meaningful representations. Experiments show that the PULSE achieves good classification performance and robustness under various lead configurations, while containing only about 0.43M parameters and enabling efficient inference on a Raspberry Pi 4B.

Critical strengths
  1. The PULSE encoder proposed in this paper has only about 0.43M parameters, offering a lightweight advantage, and it still outperforms ST-MTM, whose parameter count is 200 times larger.

  2. PULSE extracts local morphology, lead-specific morphology, rhythm, and global statistical descriptors from the complete ECG, and uses them as complementary self-supervised prediction targets to guide the model to learn signal features across different temporal scales and lead dimensions.

Critical weaknesses
  1. The baselines are insufficient and does not yet adequately validate the advantages of the proposed method. The SSL methods compared are mainly around 2021, and comparisons with recent advanced methods, such as HeartLang (Jin et al., ICLR 2025), are lacking. Second, the baselines are mainly used to compare different pretraining methods and are not sufficient to fully validate the effectiveness of the model architecture design proposed in this paper. Finally, this paper focuses on variable-lead ECG representation learning, but the selected baselines are mainly general or multi-lead ECG-oriented SSLs and are not specifically designed for variable-lead scenarios. As a method more directly relevant to this research problem, ST-MEM is currently compared only in the appendix.

  2. The analysis of robustness to missing leads is insufficient. Although PULSE achieves good absolute performance under different lead configurations, the results in Table 2 show that its performance still degrades to some extent as the number of available leads decreases, and under some settings, the magnitude of the performance degradation is greater than compared methods. Moreover, the experiments provide relatively limited coverage of different lead-missing patterns.

Questions for authors
  1. In the experiments, how were the 6-lead and 1-lead configurations determined? How did the authors establish that these combinations are representative? Can the results based on these configurations support the conclusion that the model is robust to different lead-missing patterns?

  2. In Table 2, under some settings, the performance degradation of PULSE is not better than that of the baseline methods. Do the authors consider the current experimental results sufficient to demonstrate that PULSE has an advantage in robustness to missing leads? Is the conclusion mainly based on absolute performance, or has the performance degradation caused by the reduction in leads been evaluated?

Reviewer_fAvvOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

PULSE is a compact encoder for ECGs with an arbitrary subset and ordering of the 12 standard leads. Each lead is encoded independently without a lead-identity embedding; leads then interact through a single cross-lead attention block and are pooled with symmetric operations (Cross-Lead Branch), while a parallel Deep-Sets-style branch pools per-lead summaries that do not depend on which other leads are present (Anonymous Branch). The encoder is pretrained on unlabeled 12-lead recordings (CODE-15%, Chapman-Shaoxing/Ningbo) by taking a random lead subset with temporal masking and noise and predicting fixed hand-crafted morphology and rhythm statistics of the clean signal with an InfoNCE plus Smooth-L1 objective. Downstream linear probing and fine-tuning are run on PTB-XL (5 superclasses, single-label subset), CPSC2018 and PhysioNet2017 with 12, 6 and 1 leads, against SimCLR, MoCo v3, CMSC and 3KG implemented on the same backbone. PULSE gives the best linear-probe results by a clear margin (macro-F1 higher by 0.052 to 0.309, Section 4.1) and fine-tuning results within one standard deviation of the baselines (Table 1); an appendix reports edge inference on a Raspberry Pi.

Critical strengths

1. The pretraining-objective comparison is controlled and the linear-probe gains are consistent.

Because every baseline shares PULSE's backbone and lead encoder (Section 3.2), the linear-probe gains in Table 1 (macro-F1 higher by 0.052 to 0.225 on PTB-XL and 0.062 to 0.309 on CPSC2018, as stated in Section 4.1) are attributable to the pretraining objective rather than to architecture, and they hold across three datasets and three lead configurations (Table 2).

2. The method is specified precisely enough to reproduce.

The invariance properties are stated formally (Equations 4 to 7, 10c), and the masking, lead sampling and hyperparameter details in the appendix are complete.

Critical weaknesses

I recommend clear rejection for the two reasons below. The paper's premise does not hold: lead order is known in the setting it targets, variable-lead ECG encoders are an established line of work that the paper itself cites and mischaracterizes, and neither of its two claims is tested against the natural alternative. The evaluation protocol then does not support the comparisons that are reported. The second problem is addressable with experiments; the first is the premise of the paper and would require a different paper. The Questions are offered to help a future version.

1. The premise is not valid, the setting is not new, and neither central claim is tested against its natural alternative.

Claim 1 is the lead-order-invariant set encoder for variable leads. Lead order, however, is not an unknown in the setting the paper targets. Clinical acquisition systems and the standard exchange formats (WFDB, HL7 aECG, DICOM, SCP-ECG) carry lead labels, so the order can be canonicalized before the encoder; the few legacy research databases with unnamed channels (e.g., QTDB) are two-channel Holter recordings outside this setting. Enforcing order invariance architecturally therefore buys little, while discarding lead identity throws away diagnostically relevant information, for example which leads show ST deviation.

The variable-lead setting itself is not new. Section 1 cites four approaches that already handle arbitrary lead subsets (lead masking in Oh et al., 2022; attention-based aggregation in Ruiperez-Campillo et al., 2026; contrastive distillation in Nguyen et al., 2025; arbitrary-lead multimodal supervision in Liu et al., 2025) and then asserts that "existing methods assume either a fixed lead set or need extra training per configuration". That statement is incorrect for random-lead-masking methods, which train one model for any subset, and none of the four cited methods is compared. What remains as new is lead-order invariance, which addresses the non-problem above.

Moreover, because all baselines use PULSE's set encoder (Section 3.2), the paper contains no comparison between this encoder and a conventional fixed-lead or lead-embedding encoder. The experiment that would justify the design, a variant with lead-identity embeddings under the same random-subset sampling, is missing.

Claim 2 is the "physiology-anchored" objective. Its targets are generic signal statistics (amplitude, slope, smoothness, autocorrelation) that the paper itself says are "not clinical measurements". The ablation in Table 3 removes subsets of these targets but never compares with plain masked waveform reconstruction on the same encoder, so it is unknown whether the descriptor targets help at all, and the name over-claims what the method does.

2. The evaluation protocol and the baselines do not support the comparisons.

Downstream evaluation uses random 70/10/20 five-fold splits at the level of individual 10-second inputs (Section 2.4). PTB-XL contains multiple recordings per patient (18,885 patients for 21,837 recordings; Wagner et al., 2020), and the 8,686 CPSC2018 inputs (Section 3.1) are 10-second crops of the 6,877 recordings in the public CPSC2018 set, so unless folds are formed by patient and by original recording there is leakage between training and test folds. The paper does not say, and it also does not state the filtering and normalization applied to the downstream signals.

Section 4.1 defers the only comparison with a masked-modeling ECG model, ST-MEM, to the appendix. Because ST-MEM is the most relevant competitor I checked that comparison, and it does not hold up: the ST-MEM numbers are copied from the ST-MEM paper. They were obtained on a single fixed 70/10/20 partition rather than five-fold cross-validation, with ST-MEM's own preprocessing (0.67 to 40 Hz band-pass, z-normalization), and with standard deviations over training seeds rather than folds (all from Na et al., 2024, Section 4.1 and Appendix A.2), so the two sets of numbers are not comparable. Since the downstream data construction here mirrors Na et al. (2024) almost exactly and ST-MEM's code and checkpoints are public, re-running ST-MEM under this paper's folds would have been straightforward and should be reported in Table 1. As it stands, the paper's strongest competitor is never compared on equal terms.

The remaining baselines are 2020 to 2021 contrastive methods only, and it is not stated whether they were pretrained with the same random lead dropout, which confounds the reduced-lead comparison in Table 2. The fine-tuning differences are within one standard deviation (Table 1, fine-tuning rows: PTB-XL F1 0.6786 vs 0.6764; AUROC 0.9374 vs 0.9355).

Questions for authors

These questions are meant to help a future version of the work. Q1 to Q3 would resolve my concerns about the evidence, but they would not change my recommendation, because my objection is to the premise: lead order is known, and variable-lead encoders already exist. I would reconsider only a paper reframed around the pretraining objective, with lead-order invariance dropped as a contribution, the set encoder compared with the cited variable-lead methods and with a lead-embedding variant, and the comparisons in Q1 and Q3 holding up under that framing.

  1. Please add, on the same encoder and data, (a) a variant pretrained with masked waveform reconstruction instead of descriptor targets, and (b) a variant that receives lead-identity embeddings (leads in known order) with the same random-subset sampling. These two rows would show whether the descriptor targets and the removal of lead identity are each helping or hurting.

  2. Were SimCLR, MoCo v3, CMSC and 3KG pretrained with the same random lead-subset sampling and masking as PULSE? Please state their augmentation and hyperparameter settings.

  3. Please re-run ST-MEM (public code and checkpoints) under exactly your five-fold protocol, with the same 704-d linear-probe head (Section 2.4, Equation 15) where applicable, and report it in Table 1; the copied numbers in the appendix are not comparable. Please also confirm whether the five folds are patient-disjoint on PTB-XL, or report results on the official strat_fold partition. Please also state explicitly how the downstream sets were constructed (the single-label selection of PTB-XL recordings and the 10-second cropping that yields the 8,686 CPSC2018 inputs of Section 3.1), and in particular whether the CPSC2018 folds are split by original recording or by 10-second crop; if by crop, segments of the same recording appear in both training and test folds.

  4. What exactly are the 16 morphology and 12 rhythm statistics of Section 2.3, and in what sense are they "physiological" given that Section 2.3 states they are not clinical measurements? A concrete list would let readers judge the naming; in particular, do the "rhythm" statistics include any beat-level quantity (RR interval, heart rate), or are they the same kind of statistics computed over a longer window? Please also state how the three parallel tokenizer convolutions are combined.

Clarification requests that do not affect the score: contribution 3 announces an analysis of robustness to reordered leads, but the test behind it (in the appendix) cannot fail, because both branches are permutation-invariant by construction (Equations 7 and 10c), so it should be presented as a sanity check rather than as a robustness result; the Raspberry Pi result is only in Appendix C although it headlines the abstract; Figure 1 is referenced two pages before it appears; the manuscript needs proofreading (e.g., "EXPERMENTS", "Rhytm", duplicated reference Na et al. 2024a/b).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.