acceptodds
Under review as a conference paper at ICLR 2027

ECG in Motion: Benchmarking Physiological Trajectory Modeling in Emergency Care

Abstract

Recent advances in ECG representation learning have improved static tasks such as disease classification. Continuous monitoring in emergency care, however, requires models to predict how cardiac states evolve from continuous ECG history. Existing forecasting studies largely focus on short waveform segments and rely on pointwise errors that may obscure physiological fidelity and dynamic changes. We introduce ECG-WorldBench, a benchmark built on 60,127 history-future ECG waveform pairs from 8,259 patients, with patient-disjoint splits and 10-, 30-, and 60-second models conditioned on up to 300 seconds of history. It compares periodic, deterministic, and probabilistic models using complementary waveform, physiological, distributional, and semantic measures. We also propose PSLD, a physiologically structured latent diffusion model combining a physiology-supervised latent space with multiscale historical context. Experiments reveal a key trade-off: deterministic models minimize pointwise errors, whereas periodic continuation preserves recent beat morphology but poorly follows the realized future. Among learned methods, PSLD achieves the highest waveform correlation and R-peak F1 across all horizons and the strongest distributional and semantic agreement. By exposing these complementary capabilities, ECG-WorldBench provides a unified framework for moving ECG modeling beyond static recognition toward modeling physiological evolution in continuous emergency monitoring.

Primary area:
generative models
Keywords:
ECG Generation, AI for Health, Dynamic Modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_N3RTOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces ECG-WorldBench for history-conditioned ECG waveform forecasting in emergency care. It includes 60,127 history–future pairs from 8,259 patients, forecast horizons of 10–60 seconds, and waveform, physiological, probabilistic, and representation-level evaluation. The authors also propose PSLD, combining a physiology-supervised VAE, multiscale history conditioning, a recent-waveform prefix, rhythm-phase information, and latent diffusion. The experiments highlight differences between pointwise accuracy, physiological fidelity, and temporal correspondence, with PSLD improving several metrics over the evaluated learned baselines.

Critical strengths
  1. Careful benchmark design. Continuous emergency ECG recordings, patient-disjoint splits, multiple forecast horizons, history-only normalization, and independent filtering blocks provide a useful and carefully controlled evaluation setting.

  2. Informative baseline comparisons. Periodic continuation exposes limitations of distributional and representation-level metrics. For example, Last-10s Repeat achieves better FFD and ECGFounder similarity than PSLD at the 10-second horizon despite waveform correlation of approximately 0.008 (Appendix A.6.5).

  3. Extensive empirical evaluation. PSLD achieves the lowest Energy Scores among the evaluated probabilistic models across all three horizons. Component ablations, patient-level paired tests, detector sensitivity analysis, and failure cases provide useful evidence about its strengths and limitations.

Critical weaknesses

W1. The evidence connecting PSLD's improvements to future physiological-change prediction remains incomplete.

Appendix A.12 provides encouraging evidence: PSLD achieves the best direction accuracy and magnitude error among the evaluated methods. However, change-detection recall is only 0.524–0.576, and false-positive rates reach 0.415 and 0.369 at the 30- and 60-second horizons. Direction accuracy can also count small, correctly signed predictions as successful even when they do not meet the significant-change threshold. Directional agreement therefore needs to be interpreted alongside change detection.

The change analysis evaluates a fixed generated trajectory, whereas the probabilistic results aggregate over the test population. The lower Energy Scores support better conditional forecasting overall, but do not show whether this advantage persists specifically in dynamic windows or whether predicted change probabilities agree with observed outcomes. The strong feature-space performance of periodic continuation further limits what can be inferred from FFD and representation similarity alone.

The role of longer history is similarly unresolved. At the 10-second horizon, adding the recent prefix to multiscale conditioning increases correlation from 0.105 to 0.186, while adding multiscale conditioning to a model with the prefix increases it from 0.181 to 0.186 (Appendix A.15). Longer history improves several metrics, but these ablations do not establish its contribution to predicting physiological changes. Applying the existing change metrics to the history-length and conditioning ablations would directly address this gap.

W2. The contribution needs clearer differentiation from prior work, and some secondary findings require qualification.

The distinction between local pattern reproduction and longer-range physiological dynamics is closely related to Serapio et al. (2026), “Evaluating Long-Range Temporal Structure in Foundation Model-Based Forecasts of Heartbeat Dynamics.” Broader ECG benchmarking and representation-level evaluation also appear in Tang et al. (2025), “A Comprehensive Benchmark for Electrocardiogram Time-Series.” ECG-WorldBench extends these directions through emergency-care recordings and complete-waveform forecasting over 10–60 seconds, a different target from longer-range RR dynamics. However, the new scientific insights enabled by this setting need sharper articulation. PSLD is an effective integration of established components; its empirical value is clearer than its conceptual advance.

Two secondary limitations also constrain interpretation: the 60-second R-peak F1 ranking changes under WFDB-XQRS (Appendix A.4), and clinical subgroups contain only 7–13 independent patients (Appendix A.10). These findings should remain qualified as detector-sensitive or exploratory, as the authors partly acknowledge.

Questions for authors
  1. What specific scientific insight distinguishes ECG-WorldBench from the physiological-dynamics and ECG evaluation studies discussed above?
  2. Does PSLD retain its Energy Score advantage in dynamic windows? Can multiple generated trajectories provide reliable probabilities of significant physiological changes?
  3. Could the Appendix A.12 change metrics be reported for the history-length and four conditioning configurations in Appendices A.11 and A.15?
  4. How does direction accuracy change when a prediction must also exceed the significant-change threshold?
Reviewer_TQBYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ECG-WorldBench, a benchmark for forecasting future ECG waveforms from continuous recordings in emergency care. The benchmark includes 60,127 history–future pairs from 8,259 patients, with up to 300 seconds of historical context and forecasting horizons of 10, 30, and 60 seconds. It evaluates forecasting models over diverse metrics, including waveform accuracy, physiological fidelity, probabilistic quality, and semantic consistency. The authors also propose PSLD, a diffusion-based model for conditional ECG forecasting. Experiments highlight trade-offs between deterministic and generative approaches: deterministic models achieve lower pointwise errors, while PSLD performs better on several physiological and distributional metrics among learned methods.

Critical strengths

1. A useful benchmark for continuous ECG forecasting.

The paper introduces a well-constructed benchmark with long historical contexts and multiple forecasting horizons. This provides a valuable foundation for systematically evaluating ECG forecasting models and facilitates comparisons across different forecasting paradigms. The data construction is also careful about information leakage (e.g., splits are patient-disjoint).

2. Comprehensive evaluation beyond accuracy.

The paper evaluates waveform accuracy, physiological structure, and distributional consistency, revealing meaningful trade-offs between deterministic and generative approaches. This is particularly important for physiological time series, where preserving physiologically meaningful features matters beyond point-wise accuracy. The empirical findings are convincing, which shows that lower point-wise error does not imply better physiological fidelity. The deterministic models (e.g., PSSM and PatchTST) achieve the lowest MAE/RMSE but produce waveforms with large HR and morphology errors, whereas generative model (PSLD) achieves much better physiological consistency than the other learned baselines. These findings highlight the importance of evaluating physiological forecasting models beyond conventional accuracy metrics.

Critical weaknesses

1. Insufficient motivation and unclear value of ECG waveform forecasting.

The paper does not sufficiently motivate why forecasting ECG is clinically important. Given the periodic nature of ECG signals, simple repetition of historical patterns may already produce plausible future waveforms (e.g., the Last-10s Repeat baseline achieves strong performance in Table 1; Table 6 shows that Last-10s Repeat outperforms PSLD on all distributional and semantic metrics at all horizons). Although PSLD (and some other models included in the paper) improves certain forecasting metrics, it remains unclear whether these improvements reflect meaningful prediction of physiological changes beyond periodic continuation.

Moreover, future ECG changes may be largely driven by external factors, such as patient movement or clinical interventions. It is unclear how such changes can be anticipated from historical ECG alone. The results at the 60-s horizon illustrate this concern: waveform correlation is near zero for all methods, indicating that the future waveform itself maybe essentially unpredictable at this horizon. Therefore, a clearer and stronger motivation for waveform-level forecasting would strengthen the paper.

2. Limited evidence for physiological evolution modeling.

The paper positions the benchmark as an evaluation framework for modeling physiological evolution. The history-relative change analysis in Appendix A.12 offers useful evidence beyond periodic continuation, including assessments of change direction and magnitude. However, it remains unclear to what extent the benchmark can distinguish models that learn physiological dynamics from those that simply estimate a nearby state that largely persists from recent history. From my perspective, evaluating physiological evolution should assess not only how future characteristics differ from historical ones, but also when and how these changes occur. These capabilities are related but not necessarily equivalent, and I encourage the authors to clarify this distinction and scope their claims accordingly.

3. Related Works.

The paper would benefit from a more comprehensive discussion of diffusion-based methods for time series forecasting [1]. Although NsDiff and D3U are included as baselines, Section 2 does not discuss this line of work. More importantly, for a benchmark paper, recent powerful time series foundation models such as Chronos and TimesFM are not evaluated. They have been utilized in physiological waveforms, and provedn to be powerful [2] Including them as zero-shot or fine-tuned baselines would make the benchmark more informative.

[1] Shen, L., Chen, W., & Kwok, J. (2024). Multi-Resolution Diffusion Models for Time Series Forecasting. ICLR 2024.

[2] Huang, N., Wang, H., He, Z., Zitnik, M., & Zhang, X. (2026). Repurposing Foundation Model for Generalizable Medical Time Series Classification. ICLR 2026.

Questions for authors
  1. Could you solidify why waveform-level ECG forecasting is clinically important? Specifically, what does forecasting the raw waveform offer beyond a direct prediction of physiological states and events from the historical windows?

  2. Should ECG-WorldBench be viewed as a benchmark for physiological evolution modeling? Given the concerns in Weakness 2, how does the current evaluation distinguish modeling state transitions from estimating a near-term (10/30/60 s) state that largely persists from the recent history?

I would be willing to increase my score if the authors clarify the motivation for waveform forecasting and the extent to which the benchmark evaluates physiological evolution.

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

The paper introduces ECG-WorldBench, consisting of (i) ED-ECG-Traj, a set of 60,127 history-future pairs of continuous lead-II ECG from 8,259 emergency-department patients derived from MC-MED, with patient-disjoint 90:5:5 splits, 300-second histories and 10/30/60-second futures, leakage-free block-wise filtering and history-only normalization; (ii) an evaluation protocol that combines pointwise waveform error, correlation, three peak-based "physiological" metrics (HR MAE, R-peak F1 within 50 ms, RMS-normalized median-beat template distance), probabilistic metrics (energy score, 5-95% coverage and width from K=8 samples), and distributional/semantic metrics based on frozen encoders (FFD, ECGFounder and CLEF cosine similarity, ECGFounder probability MAE), with stable/dynamic stratification defined relative to a Last-10s Repeat reference; and (iii) PSLD, a two-stage model in which a VAE with an auxiliary head predicting 26 waveform-derived descriptors provides a 10 Hz latent space and a history-conditioned diffusion transformer with multiscale history tokens, a clean 10-second prefix and rhythm/phase conditioning generates future latents. Two training-free periodic baselines, two deterministic forecasters (PatchTST, PSSM) and two latent diffusion baselines run in PSLD's VAE space are compared. Deterministic models minimize MAE/RMSE, periodic continuation best preserves HR and beat morphology, and PSLD has the best correlation, R-peak F1, energy score and semantic metrics among learned methods.

Critical strengths

1. The benchmark construction is careful about leakage and patient independence.

Section 3.1 describes patient-disjoint 90:5:5 splits, block-wise filtering that cannot leak future samples into the history, normalization estimated from the history only, and quality screening with thresholds calibrated on the training split. Section 4 further points to patient-level paired significance tests in the appendix; I checked them and they are properly done (Holm correction over the full family, patient bootstrap). The reported comparisons are therefore trustworthy as far as they go.

2. The benchmark surfaces a result the community should know: persistence of the last beats is very hard to beat.

In the main results (Table 1), Last-10s Repeat has the lowest HR MAE and morphology distance at every horizon, and Phase-Aligned Median Beat has higher R-peak F1 and correlation than PSLD at 30 and 60 seconds. The paper's own data therefore show that current learned forecasters mostly reproduce recent beat morphology. Stated plainly, this would be a valuable finding.

Critical weaknesses

I recommend clear rejection for the two reasons below. The benchmark task is mis-specified for a quasi-periodic signal, so it mostly rewards reproducing the past, and neither an argument nor an experiment connects the proposed model to a clinical decision; on the only meaningful comparison in Table 1 the training-free persistence baseline already wins. The evaluation protocol, which is the paper's second contribution, is under-defined and its metrics cannot capture the physiological changes the paper claims to target. Neither problem can be resolved by a rebuttal; the Questions are offered to help a future version.

1. The benchmark rewards reproducing the past, persistence already wins the only meaningful comparison, and nothing connects the model to a clinical decision.

The introduction itself notes that, because ECG is quasi-periodic, strong short-horizon performance "may primarily reflect the continuation of local beat patterns". The benchmark then does exactly that. By the paper's own stratification (Section 4.3), 2,005 of 2,954 test windows at 10 s are "stable", i.e., the future is indistinguishable from repeating the last 10 s, and in an ED monitoring stream most windows will be of this kind unless a clinical event occurs. Aggregate waveform, HR and morphology metrics therefore mostly measure how faithfully a method copies recent beats.

Table 1 at the 10-second horizon makes this concrete. Five of the seven methods have HR MAE above 20 bpm (Table 1, HR MAE column: Phase-Aligned Median Beat 22.531; PatchTST, PSSM and the two latent diffusion baselines between 29.839 and 30.399), which means they do not produce a plausible rhythm at all and carry no information as baselines. The comparison that remains is Last-10s Repeat against PSLD, and Last-10s Repeat is better on both HR MAE (5.712 vs 7.198 bpm) and beat morphology (0.182 vs 0.271), both from Table 1. PSLD's only advantage is R-peak F1 (0.422 vs 0.169, Table 1). That gap is an artifact of the Last-10s Repeat baseline rather than a merit of PSLD: pasting a cut segment after the history misaligns the beat timing at the splice, so its peaks fall outside the 50 ms matching tolerance (appendix metric definition) for the whole window. The training-free Phase-Aligned Median Beat removes that artifact and already reaches 0.353 (Table 1, 10 s), and at 30 and 60 s PSLD's F1 (0.231, 0.190) falls below that baseline (0.250, 0.212; Table 1). A phase-aligned repeat of the last 10 s would be a trivial baseline that likely dominates PSLD on every physiological metric, and it is not reported.

Where change does occur, the task becomes nearly unpredictable at the waveform level. Correlation with the realized future is 0.186 at 10 s, 0.052 at 30 s and 0.023 at 60 s for the best method (Table 1), so beyond the shortest horizon no method tracks the actual future, and Section 4.3 reports that PSLD degrades on every metric in the dynamic stratum. The 30- and 60-second horizons therefore do not measure forecasting; they measure how plausible a generated waveform looks.

The paper never states what a 10 to 60-second waveform forecast would be used for clinically, and no experiment ties the forecast to any decision or outcome; the clinical weak labels appear only as exploratory subgroups. A benchmark for "physiological evolution" would instead center on change: windows anchored on the recorded clinical events that MC-MED provides (and that Section 3.1 already uses for sampling), a primary target defined on derived descriptors (onset, direction and magnitude of HR, rhythm, ST or QT change), and persistence treated as the null model against which every method is scored. Since the dataset is a preprocessing and manifest layer over the public MC-MED recordings (with an unstated release mode under the PhysioNet credentialed-access terms), the data and task definition are the contribution, and they are the weakest part.

2. The evaluation protocol, contribution 2, is under-defined in the main text, and its "physiological structure" metrics cannot capture the changes the paper claims to target.

Section 3.2 gives one line per metric and leaves every definition to the appendix. A reader of the main text cannot tell what "beat-aligned morphology distance" is, and the paper never discusses what the reported values mean (Table 1, 10 s: 0.271 for PSLD versus 1.342 for PatchTST) or provides a reference scale (for example, the distance between two consecutive windows of the same patient, or between different patients). For a paper whose second contribution is the evaluation protocol, this leaves that contribution largely unsupported.

On reading the appendix definitions, the metric is the classical median beat: beats are segmented around detected peaks, a pointwise median over beats gives one representative template, and templates are compared after RMS normalization. Measuring RMSE between median beats is not an unreasonable operation, but it has known properties that make it unsuitable as the main morphology metric for "physiological evolution": it discards RR intervals and therefore says nothing about rhythm; the pointwise median suppresses minority beats, so ectopic beats such as PVCs, which are exactly the morphological changes one wants a forecaster to anticipate, are removed from the template; the RMS normalization discards amplitude; and a fixed 600 ms window around the peak (200 ms before to 400 ms after, per the appendix definition) merges P, QRS and T and overlaps neighboring beats at high heart rates. Together with HR MAE and R-peak F1, which only count peaks, the three "physiological structure" metrics contain no interval measure (PR, QRS duration, QT/QTc) and no ST or T-wave amplitude, although these are the quantities that evolve in ED patients and although the Stage-1 head already predicts QRS width.

All three metrics also depend on a peak detector that the main text does not name. The appendix describes an in-house heuristic (local maxima of the median-centered absolute amplitude above median plus 2.5 MAD, with a 300 ms refractory period) and calls its outputs "QRS proxies rather than manually annotated R peaks". Because the criterion is absolute amplitude, a large T wave or an electrode artifact can be selected instead of the QRS, which then shifts the alignment point of the median beat, and the refractory period cannot resolve rates above 200 bpm.

Every headline statement (abstract, Sections 4.1, 4.3, 5.2, the conclusion) carries the qualifier "among learned methods", and Table 2, which carries the distributional and semantic metrics behind the claim of "strongest distributional and semantic agreement", includes only the learned methods without Section 4.2 saying why. Because this omission decides the headline, I looked for the periodic baselines' values in the appendix, where they are reported separately: Last-10s Repeat beats PSLD on FFD, ECGFounder and CLEF similarity and probability MAE at all three horizons (e.g., 10 s FFD 0.0509 in the appendix table vs 0.157 for PSLD in Table 2).

Separately, the physiological supervision that gives PSLD its name is described in one sentence of Section 3.3 with no ablation in the main text; the appendix ablation shows that removing it changes downstream metrics only marginally (MAE 1.359 to 1.346 at 10 s), whereas the clean 10-second prefix accounts for most of PSLD's advantage. The model's name and contribution 3 overstate the role of physiology.

Questions for authors

These questions are meant to help a future version of the work. Answers to Q2 to Q5 can be produced from existing predictions and would make the paper more honest about its findings, but they would not change my recommendation, because my objection is to the task definition (Q1), which a rebuttal cannot resolve. I would reconsider only a paper whose benchmark is re-centered on change as described in Q1, with persistence baselines reported on equal footing throughout.

  1. Given that most windows are stable and persistence is near-optimal there, why is the benchmark not built around change? Concretely: anchor evaluation windows on the recorded clinical events in MC-MED, define the primary target as the onset, direction and magnitude of change in derived descriptors (HR, rhythm, ST level, QT/QTc), score every method as improvement over the persistence null, and report waveform generation as a secondary task. What clinical decision would the current 10 to 60-second waveform forecast inform?

  2. Please add a phase-aligned repeat of the last 10 s as a baseline (Last-10s Repeat with the boundary phase aligned as in Phase-Aligned Median Beat). Does PSLD beat it on any physiological metric? Please also explain why Phase-Aligned Median Beat, which extends a template at the estimated rhythm, has an HR MAE of 22.531 bpm at 10 s (Table 1); this suggests either the template-period estimation or the peak detector is failing on it.

  3. Please include Last-10s Repeat and Phase-Aligned Median Beat in Table 2 and in any overall ranking, and state in the abstract and conclusion that these training-free baselines are best on HR MAE, morphology, FFD and semantic similarity.

  4. Please define the physiological metrics in the main text, name the peak detector, cite R-peak F1 as the standard QRS-detection metric it is, and give a reference scale for the morphology distance (within-patient consecutive windows; between patients). Please also discuss detector sensitivity in the main text: your own appendix check with WFDB XQRS changes the R-peak F1 ranking at 60 s and reduces the number of dynamic windows from 949 to 693. Please add interval-based and amplitude-aware metrics (PR, QRS duration, QT/QTc, ST deviation, R amplitude ratio relative to history) and a metric that is sensitive to ectopic beats (e.g., per-beat morphology distance or PVC count agreement), all computable from existing predictions.

  5. The stable/dynamic criteria are thresholds on the same peak-based descriptors used as metrics, relative to the Last-10s Repeat reference, so the stable stratum is by definition where persistence works. Could you report the stratified comparison with strata defined by the recorded clinical events (or by independent annotation), and a sensitivity analysis of the six thresholds?

  6. How will ED-ECG-Traj be distributed given MC-MED's credentialed access: manifests and code only, or waveforms?

  7. Clarification, not affecting the score: does the 0.5 Hz high-pass preserve ST-segment deviations well enough for a benchmark whose motivation includes ischemic evolution? Figure 4 is too small to judge beat morphology.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.