acceptodds
Under review as a conference paper at ICLR 2027

Structure Lives in the Representation: Phase-Field Diffusion for Robust ECG Delineation

Abstract

Under noise, the masks of neural ECG delineators fragment and misorder, breaking the sequence that interval measurement presupposes. Point-level F1, the standard metric, overstates the usable fiducials. The standard remedy, rule-based post-processing, reassembles detected waves but cannot synthesize missed ones. We replace repair with generation and move structure into the learning target. We train a conditional diffusion prior over phase fields: each wave class occupies a channel whose value tracks within-wave phase. The target is continuous and invertible, encoding presence, boundaries, and order, not enforcing them afterwards. The conventional one-hot, a flat label with no within-wave position, is structurally fragile. Under the harshest external-test noise, even trained with the field's own loss, it reaches rules only by selecting among samples. Its single draw reads 0.794 against rules' 0.863. Averaging them collapses at the decision-boundary readout of established conventions. The phase field needs no such crutch. A single draw lands in rules' seed band, clearing rules at the reported pairing (0.888); its gain over the matched one-hot is positive at every seed. The same selection lifts harshest-condition legality to 0.994, above the one-hot's 0.980, and averaging stays legal without decode tuning. Legality is not bought by deleting waves: the field misses fewer ground-truth P waves than rules, attains the best F1, and trails rules on intervals by milliseconds. Under the harshest noise, generation replaces rule repair; structure steers generation.

Primary area:
learning on time series and dynamical systems
Keywords:
ECG delineation, diffusion models, phase-field representation, noise robustness, sequence legality, medical time series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Pbx6OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses structural failures in ECG delineation under noisy conditions. While point-level F1 scores for segmentation masks may appear acceptable, the outputs often contain fragmented or misordered P, QRS, and T waves, which renders downstream interval measurements unreliable.

The authors make two core contributions. First, they propose a sequence legality metric that directly evaluates whether output waveforms follow the physiological P → QRS → T order and satisfy the proposed structural ordering criterion. Second, they replace the conventional discrete one-hot mask target with a continuous phase-field representation, which embeds structural constraints into the label representation itself. A conditional diffusion model is then trained to map noisy ECG inputs to valid phase-field segmentations.

Experiments are conducted on the external QTDB test set across a 12-condition noise grid, systematically comparing a discriminative anchor model, traditional rule-based post-processing, one-hot-target diffusion, and the proposed phase-field diffusion. The paper further includes extensive ablation analyses: K-sample scans, mean aggregation behavior, wave-deletion audits, interval error quantification, multi-seed robustness, decoder control experiments, and cross-rhythm / cross-domain probes.

Critical strengths
  1. The paper raises a valid and important concern about evaluation practices in ECG delineation. Standard point-level F1 scores do not capture structural failures like fragmented or misordered waves, which directly compromise downstream interval measurements. The proposed sequence legality metric and beat-level audit complement conventional overlap-based metrics with a clinically relevant dimension.
  2. The empirical evaluation is comprehensive and well-controlled. The authors test on an external dataset across a 12-condition noise grid, and support their claims with extensive ablation studies including K-sample scans, mean aggregation behavior, wave-deletion audits, interval error quantification, multi-seed robustness, and decoder controls. This level of detail improves transparency and makes it easier to isolate the source of observed effects.
Critical weaknesses
  1. The headline 0.994 legality result is driven primarily by candidate selection rather than by the representation itself. At inference time, the method generates K=8K=8 independent candidate outputs and selects the one with the fewest illegal grammar transitions. Since legality is exactly the primary metric used to report the improvement, the inference procedure effectively searches for outputs that score well on the evaluation criterion itself.

A breakdown of the gain sources makes this clear:

  • Phase field, single draw (no selection): legality = 0.888
  • Phase field, K=8K=8 selected by legality: legality = 0.994
  • Gain from candidate selection alone: ~10.6 percentage points
  • Gain from one-hot to phase field at the same K=8K=8 selection setting: only ~1.4 percentage points

The selection gain (+10.6 pp) is substantially larger than the representation step (+1.4 pp) at the K=8K=8 operating point. This makes it difficult to attribute the headline 0.994 result primarily to the phase-field representation, and weakens the evidence for the paper's central claim that "structure lives in the representation."

  1. The legality metric has inherent limitations that are amplified by the selection mechanism. The grammar defines valid sequences as (P? QRS T)+(P?\ QRS\ T)^+, making P waves optional. This means the metric does not penalize deletion of true P waves — in fact, deleting a true P wave can produce a more legal sequence. In addition, the grammar cannot express clinically real rhythms such as non-conducted P waves in second-degree AV block.

Critically, this same flawed grammar is used to select the best of K=8K=8 candidates. The model is thus rewarded for grammar conformity even when the grammar itself does not match physiological ground truth. While the wave-deletion audit quantifies this behavior, it cannot fix the validity of the primary metric. In my view, this is a notable concern for a paper whose central claim is about structural correctness.

Questions for authors
  1. Please report results for K=1K=1 single draw, mean aggregation, and K=8K=8 legality-based selection side by side in the main comparison table. How much of the performance gain of phase field over one-hot remains when legality is not used to select among multiple samples?
  2. Could the authors use a selection rule that is completely independent of the legality metric (e.g., sorted by confidence or reconstruction error) and then report legality as an outcome? This would separate the structural robustness of the representation itself from metric-aligned search effects.
  3. How should the primary legality metric be interpreted for rhythms that cannot be expressed by the grammar (e.g., second-degree AV block)? It would be useful to report stratified results for grammar-compatible and grammar-incompatible ground-truth samples, to clarify the applicability boundary of the method across different rhythms.
Reviewer_gB7SOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies structural errors in ECG waveform segmentation under noise, including missing, fragmented, or incorrectly ordered P waves, QRS complexes, and T waves. The authors propose Phase-Field Diffusion, which uses continuous waveform representations and a diffusion model to generate segmentation predictions, aiming to reduce reliance on rule-based postprocessing. Experiments show a substantial improvement in sequence validity, along with gains in waveform recall and segmentation accuracy in some cases.

Critical strengths

The paper identifies a gap between pointwise segmentation accuracy and sequence-level structural integrity, and provides experimental evidence that this problem exists. It further introduces waveform deletion statistics and beat-level usability metrics, allowing the method to be evaluated beyond sequence validity alone. The problem has clear practical motivation and broader relevance to structured prediction.

The authors compare deterministic regression, one-hot diffusion, and continuous field diffusion. They also examine single-sample inference, candidate selection, and mean aggregation. In particular, the continuous field substantially outperforms one-hot encoding under single-sample inference. These comparisons provide fairly convincing evidence that the choice of target representation affects generation stability.

Critical weaknesses

1.High sequence validity does not necessarily imply physiologically correct predictions.

The authors evaluate predictions using a fixed P→QRS→TP \rightarrow QRS \rightarrow T order, although some arrhythmias naturally violate this pattern. Sequence validity is treated as a primary optimization and evaluation objective, but the fixed grammar cannot capture certain abnormal rhythms. Only 95.9% of the ground-truth annotations in QTDB pass the validity check, while the model achieves 99.4%. In some cases of atrioventricular block, the model produces valid outputs by deleting genuine P waves. More importantly, although the method improves beat-level measurement completeness, its complete cardiac cycle metric remains below that of rule-based postprocessing, at 0.644 versus 0.665 under the strongest noise condition. It therefore remains unclear how much of the large improvement in the primary metric translates into genuine recovery of physiological structure.

2.The necessity of the phase-field representation is not sufficiently established.

The final amplitude field is essentially a smooth waveform encoding with position-dependent gradients. It does not explicitly constrain temporal relationships among different waveform classes. Although the paper shows that it outperforms the default one-hot encoding, it does not adequately rule out alternatives such as distance fields, other smooth label encodings, or decoding with explicit structural constraints. Appendix P is particularly revealing, as the failure of mean aggregation with one-hot encoding depends heavily on the default decoding threshold. Adjusting this threshold substantially restores sequence validity. The current results therefore provide stronger evidence for the numerical stability of a particular continuous encoding, but are insufficient to establish a more general principle for structural representation.

Questions for authors

1.For arrhythmias and complete cardiac cycle evaluation, can the model genuinely improve the correctness of waveform recovery beyond increasing validity under predefined rules?

2.Why is the phase-field representation necessary? Could existing distance field encodings or other continuous label representations achieve similar single-sample stability and structural prediction performance?

Reviewer_M4TvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper elevated structural failure under noise to a first-class problem, measured it with sequence legality under a three-layer protocol, replaced rule repair with generation, and moved structure into the representation: a phase field whose single draw sits at rules’ level at the harshest cell, above the matched one-hot at every seed, whose averaging never collapses at any tested decode threshold, and whose selection reaches 0.994 above the one-hot’s 0.980. Under the harshest noise, the proposed solution seems to perform bettie than the rule-based schemes.

Critical strengths
  1. The entire paper and the proposed model look to be well motivated. The paper reads well and clear.
  2. The proposed diffusion model evidently performs better than the rule-based models.
  3. In spite of simplicity of the proposed technology, this study may encourage the audience to investigate over light-weight and small architectures to perform better rather than using large, costly and heavy deep learning models.
Critical weaknesses
  1. Whilst the proposed approach achieves better than the rule-based models by small tricks over the SOTAs in deep learning, The entire study looks trivial and does not make strong impacts on the research direction. Similar approaches have been reported in the community such as, https://arxiv.org/html/2609.16489v1 https://pmc.ncbi.nlm.nih.gov/articles/PMC11280871/ https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0330707
  2. The proposed model is called phase-field diffusion, However, in this paper, it is unclear how the phase information is utilised in the proposed model. For example, when discussing Eq. (2), how to link the angles with the physical model? by what implementation styles? any particular parameter is used to measure the phase implementation?
  3. In the experimental section, the proposed phase diffusion model does not consistently perform better than the other rule-based or one-hot diffusion approaches, e.g. Tables 2/3. This shows that the proposed model lacks robustness and generalisation ability in different conditions.
  4. The comparison methods shown in the main paper are limited. Looking at the existing similar approaches reported in the public, more comparisons are advised to take on, for example, https://proceedings.mlr.press/v219/nolin-lapalme23a/nolin-lapalme23a.pdf https://proceedings.neurips.cc/paper_files/paper/2024/file/9988f2c8e07c1f98af7ba9ca31ccae0b-Paper-Conference.pdf
Questions for authors

Please address above major concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.