acceptodds
Under review as a conference paper at ICLR 2027

READ-ECG: ECG Printout Digitisation via Structured Visual Sequence Regression for Cardiovascular Analysis

11 views

Abstract

Electrocardiograms (ECGs) are a primary cardiovascular diagnostic modality, but recordings available only as scanned or photographed printouts cannot be directly processed by signal-based AI tools. Direct image analysis entangles signal extraction with clinical interpretation, while multi-stage digitisation pipelines rely on brittle segmentation steps. We formulate ECG printout digitisation as a structured visual sequence regression problem within a decoupled Image → Signal → Analysis pipeline. We introduce READ-ECG (Reconstruction and Extraction via Attention-based Decoding of ECGs), an end-to-end framework that bypasses intermediate trace segmentation, employing a topology-aware attention-based decoder to directly regress spatial features into continuous multi-lead waveforms. To support future quality-control workflows, we incorporate aleatoric uncertainty modelling to identify potentially unreliable reconstructions. Evaluated across diverse benchmarks, READ-ECG achieves high-fidelity reconstruction on the combined PhysioNet evaluation (PCC: 0.998, SNR: 28.83 dB) and can adapt to severe real-world visual artefacts using only 81 real training images. Across three PTB-XL classification task families, a residual 1D CNN trained on original signals retains a mean AUROC of 0.927 on READ-ECG reconstructions. Evaluation with two pre-trained ECG foundation models further suggests that the digitised signals retain diagnostic information. By decoupling signal “reading” from “interpretation”, READ-ECG provides reusable waveforms to the signal-based cardiac AI ecosystem.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
ECG Digitisation, waveform reconstruction, sequence regression, medical image analysis

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 4 reviews, above 27% of the others.
Reviewer_rLTvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes READ-ECG, which models paper ECG printout digitisation as a structured visual sequence regression task. Using lead-time queries and an attention-based decoder, the method directly reconstructs multi-lead continuous waveforms from an entire ECG page, and these reconstructed signals can be processed by existing downstream ECG analysis models without retraining.

Core Claim: By exploiting the lead and temporal structure of ECGs, high-fidelity, reusable digital waveforms can be directly reconstructed from full-page images, bypassing traditional segmentation-based digitisation pipelines.

Critical strengths
  1. Unified full-page multi-lead decoding: Unlike pipelines that segment individual leads separately, READ-ECG encodes lead identity and temporal position into target queries to reconstruct multi-lead signals in a unified framework. This is a reasonable use of ECG layout structure and offers a clean implementation for full-page sequence regression.

  2. Verification of downstream signal reusability: The authors evaluate parameter-invariant downstream models on both original visible waveforms and reconstructed waveforms, showing that prediction performance barely drops. This is more meaningful than reporting reconstruction error alone, as it confirms that the digitisation preserves information utilized by existing analysis tools.

Critical weaknesses
  1. Limited conceptual novelty relative to prior work: Directly regressing waveform points from ECG images has been explored (e.g., SLPM), and the Image →\to Signal →\to Analysis pipeline is established. READ-ECG’s main addition is combining lead-time queries with full-page joint decoding. However, the paper does not show what new modeling principles this brings beyond existing image-to-sequence formulations. The method feels more like an effective engineering assembly of known techniques than a conceptual breakthrough.

  2. Lack of ablation isolating the core structured queries: Current ablations primarily examine patch size and QKNorm, without isolating the specific impact of the lead-time queries against standard sequence decoders. Thus, it is unclear whether the strong performance comes specifically from the ECG-informed query design or simply from general full-page joint regression and model capacity. In addition, evaluating different digitisation baselines under differing clean-image rendering protocols limits fair algorithmic attribution.

Questions for authors
  1. Compared to a standard image-to-sequence decoder using the same vision backbone and training data, what specific and irreplaceable advantage do the lead-time structured queries provide?
  2. Which ECG-specific structural priors are strictly necessary for the method’s performance? Can you provide evidence separating the contribution of structured query modeling from general joint sequence regression?
2 comments
AuthorsOpenReview ↗

Response to rLTv

Thank you for highlighting full-page decoding and the test of waveform reuse.

W1. Contribution relative to prior work

SLPM already regresses waveform points from individual lead crops and identifies joint full-page reconstruction as future work. Our contribution is to decode all visible leads jointly from the full printout. Each query represents a particular lead and time interval. Cross-attention lets it read the trace from the image, while self-attention lets it draw on other intervals and leads. The decoder returns a waveform and a sample-by-sample uncertainty estimate for each visible lead. SLPM

A 2024 PhysioNet Challenge entry already used digitised waveforms for classification. We ask how much performance an existing signal analyser retains when the original waveform is replaced by a reconstruction. Across three PTB-XL task families, the same residual 1D CNN scored 0.9269 mean AUROC on original visible waveforms and 0.9270 on READ-ECG outputs. It could therefore be reused on these printouts without retraining. 2024 Challenge entry

W2a, Q1–Q2. What the lead and time priors contribute

A sequence of amplitudes is usable as a multilead ECG only if each value has a lead, a time, and a voltage scale. Here, the known layout, visible lead durations, paper speed, and gain supply that mapping. Both decoders use it, so the ablation isolates how the queries represent those positions. Whether the decoder could learn the mapping from the image is a separate question.

Once each output has a lead and time, independently learned queries also reconstruct well. Factorising the queries lets the decoder reuse a lead embedding across times and a time embedding across leads. With the backbone, decoder, and training protocol held fixed, that sharing improved page SNR by 0.41 dB with 0.48M fewer trainable query parameters. The gain over frozen random queries was 1.41 dB.

We also checked how quickly the two designs learned. In one full-data fit per design, structured queries had lower normalised validation MAE (×100) than independently learned queries at epoch one (10.02 vs 18.05) and epoch ten (2.54 vs 2.77). The ordering briefly reversed at epoch two. This suggests a possible benefit with shorter training and motivates testing whether it also holds with fewer training ECGs or across different print layouts.

W2b. Same-image baselines

We agree that comparing results across renderers leaves the effect of the input style unresolved. We therefore ran released Krones M3 and Ahus Open-ECG-Digitizer on the same 2,194 clean PTB-XL pages used for READ-ECG. On 891 held-out Kaggle captures, we gave all three systems the same automatically perspective-corrected images. Ahus then applied its own preprocessing. READ-ECG used its Kaggle-adapted checkpoint, and M3 and Ahus used their released weights. Krones et al. · Ahus

MethodClean PTB-XL strictClean PTB-XL alignedKaggle strictKaggle aligned
READ-ECG30.5030.5519.4019.46
Krones M311.1914.263.699.66
Ahus11.1815.598.4012.13

READ-ECG leads on both sets of identical images, whether we score the waveforms as predicted (strict SNR) or allow a time shift of up to 0.2 s in either direction and a constant voltage offset per lead (aligned SNR). On Kaggle, alignment raises READ-ECG’s score by just 0.06 dB, compared with 5.97 dB for M3 and 3.73 dB for Ahus. We will distinguish these matched comparisons from Demolder’s and SLPM’s results on their own renderings.

Reviewer_rLTvOpenReview ↗

Good luck!

Thank you for the additional experiments. The ablation shows that structured queries improve reconstruction quality with fewer parameters. The comparisons using the same images also address my concern about the fairness of the evaluation. My remaining concern is about novelty, as the paper still does not clearly show what is new beyond existing methods. However, I think the new results are sufficient, and I will raise my rating from 2 to 3. Please include these experiments in the revised paper. Good luck!

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

The proposed READ-ECG architecture is an end-to-end framework designed to perform structured visual sequence regression directly from an ECG image. Rather than treating the image as a grid of pixels to be segmented, this paper treats it as a clinical document containing a known arrangement of signals. The proposed method has been tested over the known PhysioNet and Kaggle dataset with minor increase in the accuracy.

Critical strengths
  1. In general, the paper is understandable and not difficult to follow the proposed solution.
  2. This paper comes with sound approaches and necessary experimental work.
Critical weaknesses
  1. The proposed framework sounds like integration of standard models to deal with ECG challenges. For example, the paper states that the encoder uses a ConvNeXt-Tiny vision backbone (Liu et al., 2022), then uses a parallel, non-autoregressive Transformer decoder (Vaswani et al., 2017), whilst integrating widely-used Query-Key Normalisation (QKNorm) (Henry et al., 2020) into both the self- and cross-attention layers of the decoder. These elements and their uses are not clearly reasoned.
  2. One of the major concerns is that the motivation of the proposed research is not clear. Other published works have reported similar or better results using similar strategies, however, this paper does not clearly point out what is the critical difference between the current paper and others, just to name a few, https://pmc.ncbi.nlm.nih.gov/articles/PMC7959637/ https://www.nature.com/articles/s41598-025-07781-1 https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=11475469
  3. This paper should have explained where the proposed model fails and how to improve the systematic performance accordingly.
Questions for authors

Please address the above major concerns, where lack of novelty is the critical miss of the paper.

1 comment
AuthorsOpenReview ↗

Response to Rrbw

Thank you for asking about the design choices and failure modes.

W1. Why these components?

A sharp peak can span only a few pixels, while following the trace on either side requires a wider field of view. Fusing several stages of ConvNeXt-Tiny gives the decoder fine trace detail and the surrounding waveform context. With the same number of visual positions, multiscale fusion improved page SNR by 0.47 dB over stage-3 features alone.

Because the printout layout is known, each query already has a lead and a time interval to read. Cross-attention lets it find the corresponding trace in the page features, and self-attention lets it use context from neighbouring intervals and other leads. The decoder can therefore predict every patch in one pass, with five consecutive samples per patch. We use QKNorm to stabilise this attention during training. When we removed it, the cross-attention logits rose sharply and the loss diverged after about 25 epochs.

W2. What the method adds

The studies you cite show what can be done once a digital ECG is available. Our starting point is a scan or photograph from which we need to recover the waveform. None of the three studies evaluates that reconstruction task.

Cited studyInputTask
Naz et al.A digital ECG rendered as an imageArrhythmia classification
DeepECG-NetDigital ECG signalSignal denoising and anomaly detection
Sundaramoorthy and KarunanidhiSingle-lead digital ECGArrhythmia classification

SLPM is a close precedent for direct digitisation. It predicts waveform points from individual lead crops and identifies joint full-page reconstruction as future work. READ-ECG uses a single decoder to recover the visible waveforms jointly from the full printout and assigns uncertainty to each sample. Access to context elsewhere on the page could help refine a prediction when the local trace is ambiguous.

In a new comparison on the same 2,194 clean PTB-XL pages, strict SNR was 30.50 dB for READ-ECG, 11.19 dB for released Krones M3, and 11.18 dB for released Ahus. To see whether that fidelity makes the signal reusable, we changed only the input to the fitted residual 1D CNN. Its mean AUROC across three PTB-XL task families was 0.9269 with the original visible waveforms and 0.9270 with reconstructions.

W3. Failures and improvements

The synthetic-trained model transfers poorly to severe real artefacts. On the held-out Kaggle captures, SNR was −3.83 dB before adaptation and rose to 12.27 ± 0.22 dB across three fits after fine-tuning on 81 real training images from nine ECGs, with a separate validation set for checkpoint selection. The gain shows that training on a few representative captures helps with these artefacts. The model still assumes a fixed layout, paper speed, and gain. Learning to read calibration and layout from the page is a separate goal for future work.

Fine-tuning still leaves failures when the trace is obscured. Fig. 3b shows a spurious deflection in V2 with high uncertainty near the obscured trace. High uncertainty can help flag these reconstructions for review. For the fully adapted model on held-out Kaggle captures, setting aside the 25% most uncertain reconstructions lowered retained-set MAE from 8.50 to 6.37 µV.

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

The problem matters inasmuch as a large share of the ECGs available in pre-hospital care, in archives and in low-resource settings exist only as paper tracings or photographs, and reliable digitisation is the prerequisite for applying signal-based analysis tools to them, from delineation algorithms to ECG foundation models. The paper casts the digitisation of printed ECGs as a direct regression from the image to the twelve waveforms, without intermediate segmentation: a convolutional encoder produces a visual memory, and a Transformer decoder reads it through queries tied to a known lead-and-time layout, predicting the signal in short patches together with a per-sample aleatoric uncertainty. The layout, paper speed and gain are assumed fixed rather than read from the image.

The validation has four parts: reconstruction fidelity on synthetic renderings of several PhysioNet archives, including signals unseen in training; preservation of clinical intervals and of the performance of downstream models (a 1D CNN and two frozen foundation models) when fed reconstructed instead of original signals, contrasted with classifiers trained directly on images; the use of the predicted uncertainty to reject unreliable reconstructions; and adaptation to real photographed printouts from the Kaggle digitisation dataset, where the synthetic-trained model fails zero-shot and recovers through fine-tuning on a small number of real images.

Critical strengths

The paper is well written, technically solid and backed by a substantial experimental effort; two strengths in particular weigh in its favour.

1. An end-to-end digitiser whose output is verified to be a usable signal.\textbf{1. An end-to-end digitiser whose output is verified to be a usable signal.} Beyond SNR, clinical intervals and the performance of classifiers and frozen foundation models are preserved to within 10−310^{-3} AUROC. This is the right criterion for the stated goal (feeding the signal-based ecosystem).

2. Uncertainty that is useful for triaging reconstructions.\textbf{2. Uncertainty that is useful for triaging reconstructions.} On real images, the rejection curve tracks the oracle (MAE 6.37 vs 6.33~μ\muV at 25% rejection). The scale is calibrated only on synthetic data, but the ordering is validated on real data, which most segmentation pipelines do not provide.

Critical weaknesses

The weaknesses below are not matters of wording. They concern the evidence itself: the central claim, that direct regression improves on segmentation-based digitisation, is never tested against an existing digitiser under comparable conditions, and the downstream comparison does not measure the proposed model. I consider that added citations and a softer introduction would not address them; new experiments would.

1. No controlled comparison, neither on identical images nor on real images. Table 1 compares READ-ECG with three methods evaluated on different images, two of them through published numbers; the authors’ reproduction of ECG Digitiser (14.9 dB on their clean renderings) is below the 17.0 dB cross-validation SNR that Krones et al. report on their own ECG-Image-Kit images. On Kaggle, the only real-image test, no baseline is run, although ECG Digitiser and Open-ECG-Digitizer (Stenhede et al., npj Digit. Med., arXiv:2510.19590, Oct. 2025, not cited) are open source; the latter, trained on synthetic images only, reports 19.65 dB on real scans (10–12 dB on phone photographs), and public Kaggle solutions report about 22 dB on the competition’s hidden test set (competition metric, so not directly comparable, but indicative). READ-ECG goes from −3.8 dB zero-shot to 19.5 dB after fine-tuning on more than 7,000 real images: without a baseline under the same conditions, the abstract’s claim that existing pipelines “rely on brittle segmentation steps” which READ-ECG “bypasses” is untested.

2. The paradigm comparison does not measure READ-ECG, and the synthetic benchmark is saturated. The +0.014 AUROC gain comes entirely from a 1D CNN on clean signals versus a 2D CNN on clean images (CNN/original 0.9269 ≈ CNN/READ-ECG 0.9270); the comparison is not run on degraded images, where it would be meaningful, and the unfavourable AP on rhythm (Table A9, −0.080, interval excluding zero) is not discussed in the main text. MAEs of 1.7–3.6 µV (Table A6) at roughly 40 µV per vertical pixel (my estimate, assuming ECG-Image-Kit’s 2200×1700 renderings resized to 512×2560) indicate a benchmark that mostly measures the inversion of a known generator.

3. The introduction’s positioning is unsupported. The decoupled Image → Signal → Analysis pipeline is presented as a contribution, whereas it is the premise of the whole digitisation literature and of the PhysioNet Challenge 2024, which paired a digitisation task with a classification task that could run on the digitised waveform; no survey of the field is cited (the systematic review in J. Electrocardiol. 2023, Electronics 2026, reviews of ECG foundation models). Existing pipelines are called “brittle” although recent ones learn their stages, several estimate orientation, layout and grid scale automatically, and they reach 19.65 dB on real scans without any real paired supervision (Stenhede et al.), or about 22 dB on the Kaggle hidden test set with the same real training data as the authors (public Kaggle solutions); READ-ECG instead hard-codes the geometry (layout, 25 mm/s, 10 mm/mV) and delegates perspective correction to an undescribed step. The reconstructed signal is called “artefact-free” although the authors’ own Fig. 3b shows a hallucinated QRS. Finally, the premise about the strength of ECG foundation models is at least qualified by Table A7, where frozen ECG-JEPA and ST-MEM with MLP probes score below the from-scratch 1D CNN (a frozen-probe setting, admittedly, but the one the authors chose).

Questions for authors

Clarification requests

  1. Is the “standard perspective correction” applied to the Kaggle images automatic? Is the 19.5 dB obtained without any manual input (corners, orientation, calibration)?

  2. Which model produces the PTB-XL row of Table 1: the MAE-trained or the Laplace-NLL-trained one? Does the uncertainty objective change reconstruction fidelity, in particular on QRS peaks? (A heteroscedastic NLL can inflate b on hard regions instead of fitting them; Seitzer et al., ICLR 2022.)

  3. Why is your reproduction of ECG Digitiser (14.9 dB on clean renderings) below the 17.0 dB cross-validation SNR that Krones et al. report on augmented images? Which checkpoint, training data and preprocessing were used?

Additional evidence that could affect my assessment.

  1. At least one existing digitiser (ECG Digitiser, Open-ECG-Digitizer, or a public Kaggle solution) evaluated on your Kaggle test split with your SNR definition; and, symmetrically, READ-ECG on PM-ECG-ID (public) or PaperECG_12l (if available), so that one comparison is made on identical images.

  2. An ablation of the topology-aware queries (against free learned queries) and of the multi-scale fusion, the two components presented as specific to the method.

  3. The paradigm comparison on degraded images (Kaggle if labels exist, otherwise ECG-Image-Kit with its full distortions), with a comment on the unfavourable rhythm AP of Table A9.

  4. Uncertainty calibration on the Kaggle test set (coverage or AUCE), conditional on degradation level and on QRS peaks, with a sharpness measure; rejection curves alone only test the ordering.

Corrections and suggestions (no effect on the score).

  1. Barros et al. 2024 is a signal-based study; the citation as an image-to-diagnosis example (§1, §2) should I believe be corrected. Reconcile 19.40 / 19.48 / 19.5 dB for Kaggle.

  2. Future work: reading printed calibration and layout from the image; amplitude-based morphology (ST segment, P and T waves) against an independent reference such as PTB-XL+.

3 comments
AuthorsOpenReview ↗

Response to jFgG (1/3)

Thank you for the detailed review and for highlighting waveform reuse and uncertainty.

W1a, Q3. Clean PTB-XL comparison

The 17.02 dB you cite is M1’s published cross-validation score. Our Table 1 outputs closely match released M3, which scored 15.27 dB on augmented renderings. In the M1 examples we inspected, segmentation consistently missed sharp QRS peaks in lead II, so we focused on M3.

For the new comparison, we used the released M3 checkpoint and its rotation, segmentation, and vectorisation pipeline. Krones et al. trained M3 on 69,000 synthetic images and fine-tuned it on 100 Challenge images.

Our M3 exporter inadvertently compressed the 10-second lead II strip to 2.5 seconds. Omitting that strip and standardising each lead before alignment reproduces the 14.94 dB reported in Table 1. After correcting the exporter, we evaluated all 12 leads on the same 2,194 clean PTB-XL pages using the amplitude-preserving alignment applied to every method below. M3 scored 14.26 dB, about 1 dB below its published cross-validation result. We will correct Table 1 and label the published scores separately. All scores are in dB. Krones et al.

MethodStrictAlignedChallenge-styleAligned gainPublished CV
READ-ECG30.5030.5531.18+0.05-
Krones M12.362.811.95+0.4517.02
Krones M311.1914.2613.85+3.0715.27
Ahus11.1815.5915.91+4.41-

Allowing a time shift and a baseline offset improves SNR by just 0.05 dB for READ-ECG, compared with 3.07 dB for M3 and 4.41 dB for Ahus. Predicting the amplitude directly at each time point may help READ-ECG preserve the trace’s timing and baseline.

W1b, W3a, Q4. Kaggle baselines and SNR

Following your suggestion, we evaluated released M3 and Ahus Open-ECG-Digitizer on the same 891 Kaggle test captures (99 ECGs). Kaggle-adapted READ-ECG leads both released digitisers under all three SNR definitions. For the first three rows, all systems received the same images after automated perspective correction (§3.2). Ahus still applied its own preprocessing. We also ran Ahus on the original images to check whether our correction interfered with its native pipeline.

MethodStrictAlignedChallenge-styleAligned gain
READ-ECG, Kaggle-adapted19.4019.4620.13+0.06
M3, released3.699.669.35+5.97
Ahus, released8.4012.1313.00+3.73
Ahus, released (native preprocessing)2.398.4510.21+6.06

On the shared inputs, alignment adds just 0.06 dB for READ-ECG, compared with 5.97 dB for M3 and 3.73 dB for Ahus. We will cite Ahus and distinguish the three scores in §3.4.

  1. Strict SNR. We compute 10log⁡10∑tst2∑t(st−s^t)210\log_{10}\frac{\sum_t s_t^2}{\sum_t(s_t-\hat{s}_t)^2} on the visible samples of each lead, where sts_t is the reference and s^t\hat{s}_t the prediction. We average the dB scores across leads and images.
  2. Aligned SNR. We allow a time shift of up to 0.2 s in either direction and a constant voltage offset per lead, then average as above.
  3. Challenge-style SNR. We pool aligned signal and error energies across leads within each image, then average the resulting power ratios across images before converting to dB.
AuthorsOpenReview ↗

Response to jFgG (2/3)

W2a, W3e, Q6. Classification

You are right that the 1D CNN already leads ConvNeXt-Tiny on original PTB-XL waveforms (0.9269 vs 0.9129 mean AUROC). READ-ECG preserves that advantage, with 0.9270 for the same 1D CNN on reconstructions.

The same comparison shows a deficit in rhythm AP (−0.080, 95% CI [−0.124, −0.022]). The mean AP difference is smaller and its interval includes zero (−0.010 [−0.027, +0.012]). The rhythm gap is already present with the 1D CNN on original waveforms (0.578 vs 0.660) and is concentrated in atrial flutter, PSVT, and trigeminy. With only 3, 1, and 3 positives, a few rank changes can shift macro-AP.

To test both approaches on degraded images, we evaluated 810 captures from 90 labelled Kaggle test ECGs. ConvNeXt-Tiny and READ-ECG received the same perspective-corrected images. Each signal model kept its fitted weights as we switched its input from the original waveform to the reconstruction. These patients were held out during classifier fitting and selection.

Model or training routeMean AUROCMean AP (pooled)
Fixed signal models
Original waveform → 1D CNN0.91990.7709
READ-ECG → same 1D CNN0.92010.7515
Original waveform → ECG-JEPA0.92850.8204
READ-ECG → ECG-JEPA0.92940.8110
Original waveform → ST-MEM0.93620.7741
READ-ECG → ST-MEM0.93340.7265
Image classifiers
ConvNeXt-Tiny (ImageNet init → Kaggle)0.86060.6084
ConvNeXt-Tiny (ImageNet init → clean PTB-XL → Kaggle)0.90650.7322
ConvNeXt-Tiny (adapted READ-ECG encoder init → Kaggle)0.77090.4519
ConvNeXt-Tiny (adapted READ-ECG encoder init → clean PTB-XL → Kaggle)0.91750.7269

The fixed 1D CNN retained nearly the same AUROC on reconstructions, although pooled AP fell. Even when ConvNeXt-Tiny started from the adapted READ-ECG encoder and trained on clean PTB-XL before Kaggle, it reached 0.9175 AUROC and 0.7269 pooled AP. On those same captures, READ-ECG + 1D CNN scored higher on both metrics (0.9201 and 0.7515), as did READ-ECG + ECG-JEPA (0.9294 and 0.8110). Both paired 95% intervals for the 1D CNN comparison with this ConvNeXt-Tiny include zero.

On clean PTB-XL, the 1D CNN led both frozen probes in AUROC. On degraded Kaggle reconstructions, both frozen probes scored higher than the 1D CNN in AUROC, and ECG-JEPA had the highest pooled AP (0.8110, or 0.8212 averaged across capture types). Recovering the waveform thus lets us apply these pretrained signal models directly to ECGs available only as printouts.

W2b, W3b. Scope and contribution

We deliberately use clean renderings to measure how accurately the decoder recovers a waveform in a known layout. A rendered trace can carry position cues between pixel centres, so mean amplitude error can be smaller than the amplitude represented by one vertical pixel. We use the resulting model to initialise real-image adaptation. Before adaptation, it reaches 26.49 dB on controlled MIMIC-IV renderings but only −3.83 dB on Kaggle. The clean experiment therefore tells us how accurately the decoder can read a known format. The Kaggle experiment tests how much real-image training it needs to handle capture artefacts.

As you point out, a 2024 Challenge entry already used reconstructed waveforms for classification. With trace segmentation, steep deflections can occupy several pixels in the same column, and peaks can be lost when converting the mask to a waveform, as Krones et al. observed. READ-ECG bypasses that step by jointly regressing waveform samples and their uncertainties across the visible leads from the full printout. We then feed the reconstructions to the same fitted signal models used on original waveforms to test whether existing analysers can use them. We will revise the introduction to credit the established workflow, cite the surveys, and make this contribution explicit. 2024 Challenge entry · Krones et al. · 2023 review · 2026 survey

AuthorsOpenReview ↗

Response to jFgG (3/3)

W3c–d, Q1. Geometry and failure modes

Yes, the reported Kaggle result uses automated perspective correction. The model is trained for a fixed 3×4 layout plus a rhythm strip, with 25 mm/s paper speed and 10 mm/mV gain.

Inferring geometry is future work. Varying speed and gain in training could let the decoder read printed settings and the calibration pulse and apply the scales automatically. For variable layouts, we would also enlarge the query bank to cover a 12×1 printout and predict both presence and waveform values for each lead–time slot.

Fig. 3b shows a spurious deflection in an obscured trace. We will remove the “artefact-free” wording.

Q2. Training objective and QRS fidelity

The paper’s main READ-ECG reconstruction results, including Table 1’s PTB-XL result, use Laplace-NLL checkpoints. To test the effect on peaks, we froze the visual features and continued training the same checkpoint for five epochs with either MAE or Laplace NLL. On lead II, switching to MAE reduced the average error within QRS windows from 7.79 to 7.30 µV and the error in peak height relative to the local baseline from 17.81 to 14.73 µV. This short comparison supports your concern that learning uncertainty can come at a cost around sharp peaks. Reducing that error while retaining uncertainty estimates is a goal for future work.

Q5. Query and fusion ablations

Giving queries explicit lead and time positions improved page SNR by 0.41 dB over independently learned queries and 1.41 dB over frozen random queries. The structured design also uses 0.48M fewer trainable query parameters than the independently learned design. With the same number of image features, multiscale fusion added 0.47 dB over stage-3 features alone.

We also checked how quickly the two designs learned. In one full-data fit per design, structured queries had lower normalised validation MAE (×100) than independently learned queries at epoch one (10.02 vs 18.05) and epoch ten (2.54 vs 2.77). The ordering briefly reversed at epoch two. This suggests a possible benefit with shorter training and motivates testing with fewer training ECGs.

Q7. Real-image calibration

As you note, uncertainty-based rejection comes close to the oracle on Kaggle (6.37 vs 6.33 µV retained MAE at 25% rejection). To test whether the uncertainty also quantifies the size of the error, we fitted a single scale multiplier on validation data. On the test set, it reduces average calibration error (AUCE) from 0.122 to 0.039. We report empirical coverage and mean total width for nominal 95% intervals. QRS windows come from the reference waveforms.

Stratum95% Coverage (raw → scaled)95% Width, raw → scaled (µV)
All91.07% → 97.79%35.98 → 50.90
Reference QRS89.64% → 97.03%163.67 → 231.58
Paper photo90.54% → 97.68%33.71 → 47.69
Mould-damaged greyscale scan89.82% → 97.16%41.28 → 58.41

The reduction in average calibration error comes with wider intervals. Across the nine capture types, coverage at nominal 95% rises from 89.82–94.53% to 97.16–98.97%, so the scaled intervals overcover at that level. Multiplying all scales by the same factor leaves the rejection ranking unchanged.

Q8. Citation and Kaggle scores

Thank you for catching the Barros citation – we will remove it as an example of printout classification in §§1–2. We will also update Table 1’s Kaggle SNR from the single-fit result of 19.40 dB to the three-fit mean of 19.48 dB, which rounds to 19.5 dB in the text.

Q9. Future work on geometry and morphology

Beyond the geometry extensions in Q1, we would test ST-segment, P-wave, and T-wave amplitudes against an independent reference such as PTB-XL+.

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

Digitises ECG printouts by regressing all 12 lead waveforms directly from the page image. A ConvNeXt-Tiny encoder feeds a 3-layer decoder whose queries carry lead identity and time position; each query outputs 5 samples plus a Laplace scale for uncertainty. Trained on ECG-Image-Kit renderings. Reports 30.5 dB on clean PTB-XL, 19.4 dB on Kaggle photos after fine-tuning (−3.8 dB zero-shot), interval agreement, uncertainty-based rejection, and a downstream check where a 1D CNN and two frozen foundation models are fed the reconstructed signals.

Critical strengths
  1. Whole-page regression with no cropping or segmentation is a clean design. The QKNorm stability analysis and the per-sample uncertainty (r = 0.81 with error; rejection curve close to oracle) are solid, and code plus demo are provided.
  2. Honest about the real-image gap: zero-shot −3.8 dB, failure cases, and the few-shot curve are all reported, and downstream comparisons come with paired bootstrap intervals.
Critical weaknesses
  1. The paper’s thesis is that direct regression is more robust than segmentation pipelines on degraded images, but that is never tested on degraded images. The four rows of Table 1 come from four different test sets; only ECG Digitiser is reproduced, and the reproduction (14.94 dB) is below what Krones et al. report on their own augmented synthetic images (17.02 dB). On the Kaggle photos there is no baseline at all, even though competition solutions on this very dataset and the open-source npj 2025 pipeline are available and uncited.
  2. The paradigm comparison is run on clean synthetic renderings, which does not match the stated motivation about real artefacts. Even there it wins only on AUROC; AP is −0.010 overall and −0.080 on rhythm (Table A9), not mentioned in the text. Also, the paper never states whether the 977 Kaggle PTB-XL records were excluded from the synthetic training set, so the few-shot results may have signal-level leakage.
Questions for authors
  1. Can you run at least one baseline on your Kaggle split (the ECG Digitiser reproduction you already have would do), or submit to the official test set? Please also define SNR in the main text.
  2. Do the 977 Kaggle PTB-XL records overlap with the 18,527 PTB-XL records used for synthetic training?
  3. The 1% fine-tuning selects checkpoints on a 432-image real validation set. Does the result hold if selection uses only the 81 training images?
2 comments
AuthorsOpenReview ↗

Response to MTyw (1/2)

Thank you for your suggestions on the baseline comparisons, downstream performance, and few-shot evaluation.

W1a. Clean PTB-XL comparison

The 17.02 dB you cite is M1’s published cross-validation score. Our Table 1 outputs closely match released M3, which scored 15.27 dB on augmented renderings. In the M1 examples we inspected, segmentation consistently missed sharp QRS peaks in lead II, so we focused on M3.

Our M3 exporter inadvertently compressed the 10-second lead II strip to 2.5 seconds. Omitting that strip and standardising each lead before alignment reproduces the 14.94 dB reported in Table 1. After correcting the exporter, we evaluated all 12 leads on the same 2,194 clean PTB-XL pages using the amplitude-preserving alignment applied to every method below. M3 scored 14.26 dB, about 1 dB below its published cross-validation result. We will correct Table 1 and label the published scores separately. All scores are in dB. Krones et al.

MethodStrictAlignedChallenge-styleAligned gainPublished CV
READ-ECG30.5030.5531.18+0.05-
Krones M12.362.811.95+0.4517.02
Krones M311.1914.2613.85+3.0715.27
Ahus11.1815.5915.91+4.41-

Allowing a time shift and a baseline offset improves SNR by just 0.05 dB for READ-ECG, compared with 3.07 dB for M3 and 4.41 dB for Ahus. Predicting the amplitude directly at each time point may help READ-ECG preserve the trace’s timing and baseline.

W1b, Q1. Kaggle baselines and SNR

Following your suggestion, we evaluated released M3 and Ahus Open-ECG-Digitizer on the same 891 Kaggle test captures (99 ECGs). Kaggle-adapted READ-ECG leads both released digitisers under all three SNR definitions. For the first three rows, all systems received the same images after automated perspective correction (§3.2). Ahus still applied its own preprocessing. We also ran Ahus on the original images to check whether our correction interfered with its native pipeline.

MethodStrictAlignedChallenge-styleAligned gain
READ-ECG, Kaggle-adapted19.4019.4620.13+0.06
M3, released3.699.669.35+5.97
Ahus, released8.4012.1313.00+3.73
Ahus, released (native preprocessing)2.398.4510.21+6.06

On the shared inputs, alignment adds just 0.06 dB for READ-ECG, compared with 5.97 dB for M3 and 3.73 dB for Ahus. We will cite Ahus and distinguish the three scores in §3.4.

  1. Strict SNR. We compute 10log⁡10∑tst2∑t(st−s^t)210\log_{10}\frac{\sum_t s_t^2}{\sum_t(s_t-\hat{s}_t)^2} on the visible samples of each lead, where sts_t is the reference and s^t\hat{s}_t the prediction. We average the dB scores across leads and images.
  2. Aligned SNR. We allow a time shift of up to 0.2 s in either direction and a constant voltage offset per lead, then average as above.
  3. Challenge-style SNR. We pool aligned signal and error energies across leads within each image, then average the resulting power ratios across images before converting to dB.
AuthorsOpenReview ↗

Response to MTyw (2/2)

W2a. Downstream classification

As you noted, on clean renderings READ-ECG + 1D CNN gains 0.014 mean AUROC over ConvNeXt-Tiny, with a mean AP difference of −0.010 (95% CI [−0.027, +0.012]) and a rhythm AP difference of −0.080 ([−0.124, −0.022]). The rhythm gap is already present on the original waveforms. The 1D CNN scores 0.578 versus 0.660 for ConvNeXt-Tiny, and digitisation raises the 1D CNN’s score slightly to 0.580. The gap is concentrated in atrial flutter, PSVT, and trigeminy. With only three, one, and three positive test ECGs, respectively, a few rank changes can shift macro-AP.

To test both approaches on degraded images, we evaluated 810 captures from 90 labelled Kaggle test ECGs. ConvNeXt-Tiny and READ-ECG received the same perspective-corrected images. Each signal model kept its fitted weights as we switched its input from the original waveform to the reconstruction. These patients were held out during classifier fitting and selection.

Model or training routeMean AUROCMean AP (pooled)
Fixed signal models
Original waveform → 1D CNN0.91990.7709
READ-ECG → same 1D CNN0.92010.7515
Original waveform → ECG-JEPA0.92850.8204
READ-ECG → ECG-JEPA0.92940.8110
Original waveform → ST-MEM0.93620.7741
READ-ECG → ST-MEM0.93340.7265
Image classifiers
ConvNeXt-Tiny (ImageNet init → Kaggle)0.86060.6084
ConvNeXt-Tiny (ImageNet init → clean PTB-XL → Kaggle)0.90650.7322
ConvNeXt-Tiny (adapted READ-ECG encoder init → Kaggle)0.77090.4519
ConvNeXt-Tiny (adapted READ-ECG encoder init → clean PTB-XL → Kaggle)0.91750.7269

The fixed 1D CNN retained nearly the same AUROC on reconstructions, although pooled AP fell. Even when ConvNeXt-Tiny started from the adapted READ-ECG encoder and trained on clean PTB-XL before Kaggle, it reached 0.9175 AUROC and 0.7269 pooled AP. On those same captures, READ-ECG + 1D CNN scored higher on both metrics (0.9201 and 0.7515), as did READ-ECG + ECG-JEPA (0.9294 and 0.8110). Both paired 95% intervals for the 1D CNN comparison with this ConvNeXt-Tiny include zero.

W2b, Q2. Source overlap

Yes. Waveform matching linked 921 of 977 Kaggle records to PTB-XL. Of the 99 Kaggle test ECGs, 76 matched the synthetic training set and six matched synthetic validation. In the full-data fine-tuning runs we audited, the synthetic training images also included renderings of nine test ECGs.

Full retraining exceeded our rebuttal compute budget, so we tested what extra exposure to known waveforms would change. We continued training twice from the same synthetic-trained checkpoint (PhysioNet, Fig. 6). In one run, clean renderings of 32 test ECGs replaced 32 synthetic controls. Both runs then fine-tuned on the same 81 real Kaggle images. On those 32 ECGs, extra exposure raised strict SNR by just 0.13 dB (95% CI [0.02, 0.24]) and changed MAE by −0.02 µV ([−0.19, 0.16]). Across all 99 test ECGs, both runs gained roughly 15 dB after fine-tuning. This points to few-shot adaptation as a major contributor to performance on new captures in this cohort. We expect adaptation to remain beneficial when the source ECGs are held out entirely.

Q3. Checkpoint selection

Yes. Few-shot adaptation also improved reconstruction when we fixed the training duration in advance. Each of the three fits used 81 real training images and synthetic images. We evaluated fixed epochs without real validation for checkpoint selection, then used separate validation subsets to test how much the result depended on the size of the validation set.

Checkpoint ruleReal validation imagesStrict SNR (dB)
Before fine-tuning0−3.83
Fixed epoch 50011.86 ± 0.16
Fixed epoch 100011.79 ± 0.11
Fixed epoch 200011.49 ± 0.09
Lowest MAE (single image)112.02 ± 0.11
Lowest MAE (single ECG, 9 captures)912.04 ± 0.14
Lowest MAE (full validation split)43212.17 ± 0.17

Values are mean ± SD across fits. Averaged over draws, selecting with a single validation image gave 12.02 dB, compared with 12.17 dB using all 432 images.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.