Pith. sign in

REVIEW 3 major objections 5 minor 34 references

ImputeECG reconstructs complete 12-lead, 10-second ECGs from incomplete recordings while keeping every observed sample, restoring diagnostic AI performance close to complete-ECG levels.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-07-11 10:10 UTC pith:FVIKU75E

load-bearing objection Solid applied ECG completion paper: real gains under clean simulated masks and external diagnostics, but the real-world AI-ready claim still rests on unpaired sex/age lifts after digitization. the 3 major comments →

arxiv 2607.05009 v1 pith:FVIKU75E submitted 2026-07-06 cs.LG cs.AI

ImputeECG: Deep Learning Reconstruction of Complete 12-Lead Electrocardiograms from Incomplete Recordings for Cardiac Assessment

Xiaocheng Fang , Haoyu Wang , Jieyi Cai , Qinghao Zhao , Jun Li , Shanwei Zhang , Guangkun Nie , Yujie Xiao
show 9 more authors
Shun Huang Jiarui Jin Hongmin Liu Guodong Wang Shuohua Chen Liming Lin Shouling Wu Hongyan Li Shenda Hong
This is my paper
classification cs.LG cs.AI
keywords 12-lead ElectrocardiographyIncomplete ECG RecordsECG Signal CompletionMask-conditioned TransformerAI-enabled Cardiac AssessmentECG digitizationTransformer autoencoder
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Many clinical ECGs remain incomplete after paper or image digitization because only short lead segments are displayed or because parts of the waveform are lost or corrupted. The paper presents ImputeECG, a mask-conditioned one-dimensional Transformer autoencoder that fills missing regions of a 12-lead, 10-second ECG while leaving all observed samples unchanged. Trained on PTB-XL and tested on PTB-XL, external CPSC2018, and a 43,633-record real-world digitized hospital cohort, it cuts missing-region reconstruction error roughly in half versus prior methods, better preserves timing and wave morphology, and restores multi-label diagnostic and demographic prediction scores toward the performance of complete digital ECGs. A reader who cares about AI-ready cardiac data would see this as a practical way to unlock incomplete ECG archives without discarding the original observed signal.

Core claim

ImputeECG, a mask-conditioned one-dimensional Transformer autoencoder, completes incomplete 12-lead, 10-second ECGs while retaining all observed samples exactly. On PTB-XL it reduced missing-region MAE by 41.7–51.0% and MSE by 54.0–63.7% versus the strongest baseline, with lower errors in R-peak timing, RR, QRS, QT, and P/QRS/T morphology; on CPSC2018 MAE fell 49.7–51.9%. Downstream multi-label classification approached complete-ECG AUROC/AUPRC, and after image digitization in a large clinical cohort it improved zero-shot sex prediction AUROC and age prediction MAE.

What carries the argument

Mask-conditioned 1D Transformer autoencoder: the zero-filled observed ECG is concatenated with a binary missingness mask (24 channels), patch-embedded, encoded by a 12-block Transformer that models long-range time and cross-lead structure, then decoded to reconstruct missing patches; at inference the completed signal is observed samples kept exactly plus model fill-in only where the mask marks missing.

Load-bearing premise

The main evidence assumes that simulated short-display layouts and random temporal gaps adequately stand in for real-world incompleteness from paper and image digitization, including grid artifacts, scanning distortion, trace overlap, and unrecovered segments.

What would settle it

On paper or PDF ECGs that also have paired complete digital reference waveforms, check whether missing-region MAE and downstream diagnostic AUROC stay within the margins reported for the simulated 4×3 and 6×2 settings; a large gap would show the method does not transfer to true digitization incompleteness.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Incomplete digitized paper or short-display ECGs can be turned into standardized 12-lead, 10-second digital signals for AI models.
  • Historical ECG archives that lack full digital waveforms become usable without re-recording patients.
  • Classifiers trained on complete ECGs can be applied to completed records with performance near the complete-ECG reference.
  • Real-world short-display and partially corrupted digitized ECGs gain measurable utility for zero-shot clinical prediction after completion.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same mask-plus-Transformer pattern could transfer to other multi-channel biosignals with short-display or sensor-dropout missingness, such as multi-lead EEG or wearable PPG arrays.
  • Explicit uncertainty over imputed regions would be a natural clinical safety layer when rare events sit entirely inside missing segments.
  • Tight coupling of digitization and completion in one pipeline could reduce separate quality-control steps on partially recovered traces.
  • If morphology holds under true clinical noise, retrospective multi-site ECG-AI studies could expand by including image-only archives.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. ImputeECG is a mask-conditioned 1D Transformer autoencoder that reconstructs complete 12-lead, 10-s ECGs from incomplete recordings while exactly retaining all observed samples (Eqs. 1–11). Trained on PTB-XL under three simulated missingness patterns (4×3 2.5 s, 6×2 5 s, and local temporal gaps with random continuous-gap augmentation), it is evaluated on PTB-XL and external CPSC2018 for missing-region waveform fidelity (MAE/MSE/FD/FID/SSIM/PSNR/ACD/MDD), morphology (R-peak, RR, QRS, QT, P/QRS/T regions), and fixed Net1D multi-label diagnostic AUROC/AUPRC, plus a 43,633-record Kailuan real-world cohort of image-digitized short-display ECGs for zero-shot sex and age prediction with a frozen HEEDB-pretrained Net1D. On PTB-XL the method reduces missing-region MAE by 41.7–51.0% and MSE by 54.0–63.7% versus the strongest baseline and restores AUROC/AUPRC near complete-ECG levels (e.g., 92.28%/33.88% in the 4×3 setting); similar gains hold on CPSC2018, and Kailuan shows modest zero-shot improvements (sex AUROC 82.6%→85.8%; age MAE 10.72→9.87 y).

Significance. If the results hold under realistic digitization noise, the work supplies a practical, mask-aware completion step that can convert short-display and partially recovered clinical ECG archives into standardized 12-lead 10-s signals usable by existing ECG-AI models. Strengths include subject-level splits, metrics restricted to originally missing regions, independent fixed downstream classifiers with bootstrap CIs, external CPSC2018 transfer, public code, and an honest Discussion of limitations. The architecture is a natural 1D MAE-style adaptation rather than a fundamental methodological breakthrough, but the clinical framing and multi-level evaluation (waveform, morphology, diagnostic utility, real-world zero-shot) make the contribution useful for digital cardiovascular medicine.

major comments (3)
  1. The central claim that ImputeECG converts incomplete clinical (especially image-digitized) records into AI-ready complete ECGs rests almost entirely on clean simulated masks applied to already-digital complete waveforms (Methods, Eqs. 9–11; Tables 1–3; Figs. 2–3). Real digitization incompleteness—grid artifacts, scanning distortion, trace overlap, baseline drift, low-resolution boundaries, imperfect digitizer recovery—is acknowledged in the Discussion but never used for paired waveform-level evaluation. Kailuan supplies only unpaired zero-shot sex/age gains with no complete digital references, so waveform fidelity and disease-feature recovery under true digitization noise remain unmeasured. This gap is load-bearing for the real-world claim and should be narrowed (e.g., by synthesizing realistic digitization artifacts on PTB-XL/CPSC, or by reporting any available paired digital/image subs
  2. Downstream diagnostic utility on PTB-XL and CPSC2018 is assessed with fixed Net1D multi-label classifiers, which is methodologically clean, but the Kailuan tasks (sex AUROC +3.24 pp; age MAE −0.85 y) are weak proxies for cardiac diagnostic fidelity. Sex and age can improve from any amplitude completion or regularization that helps a pretrained Net1D without proving that reconstructed P/QRS/T morphology or transient abnormalities are clinically faithful. The manuscript should either add disease-relevant labels (or at least rhythm/conduction proxies) on Kailuan if available, or explicitly frame Kailuan as usability evidence only and not as clinical diagnostic validation.
  3. Discussion correctly notes that transient abnormalities, premature beats, ischemic changes, or pacing artifacts lying entirely inside missing regions may be unrecoverable from surrounding context, yet no quantitative stress test of this failure mode is provided. A stratified analysis (e.g., reconstruction and diagnostic error conditioned on presence of PVCs/AF/ischemia inside vs. outside the mask) would substantially strengthen the safety interpretation of the restored AUROC/AUPRC numbers in Tables 1 and 3.
minor comments (5)
  1. Figure 1 and the later architecture figure (Fig. 6) contain duplicated panels and inconsistent scenario labels (e.g., “6×2-lead, 10s” vs. “12×1”); clean for publication.
  2. PTB-XL record counts differ slightly between abstract/intro (21,799) and Methods figure text (21,837); reconcile.
  3. Preprocessing is inconsistent across cohorts (no z-score on PTB-XL; per-segment z-score on CPSC2018; per-lead z-score on Kailuan). A short justification or sensitivity check would help readers assess transfer.
  4. Baselines (CycleGAN, Pix2Pix, EKGAN, ECGRecover) are appropriate, but brief notes on hyperparameter matching and whether they also receive the exact binary mask would improve fairness claims.
  5. Code and public-data links are provided; ensure the released repository includes the exact mask-generation scripts and evaluation configs used for Tables 1–3 so the missing-region metrics are fully reproducible.

Circularity Check

0 steps flagged

No significant circularity: empirical ML reconstruction evaluated on held-out public benchmarks and independent downstream models, with no self-definitional equations or fitted-as-prediction reductions.

full rationale

ImputeECG is a standard mask-conditioned 1D Transformer autoencoder trained with L1 loss only on missing samples (Eq. 10) under simulated incomplete patterns (Eq. 9) and evaluated by comparing completed signals to held-out complete references on PTB-XL and external CPSC2018 (waveform MAE/MSE/SSIM etc. over missing regions only; morphology timing errors; multi-label AUROC/AUPRC of independently trained Net1D classifiers). Inference retains observed samples exactly (Eq. 11). Kailuan uses a frozen HEEDB-pretrained Net1D for zero-shot sex/age without fine-tuning the imputer. No equation equates a claimed prediction to a fitted free parameter by construction; no uniqueness theorem or ansatz is imported from overlapping-author citations as a load-bearing premise; self-citations (related ECG works) are non-essential background. The derivation chain is ordinary supervised training plus external evaluation and is self-contained against public benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central empirical claim rests on standard deep-learning practice plus domain assumptions about 12-lead physiological redundancy; free parameters are ordinary architecture and optimization choices; the only invented entity is the named model itself, which is fully specified and open-sourced.

free parameters (4)
  • patch size P = 50
    Fixed at 50 samples (0.1 s at 500 Hz), determining the 100-token sequence length; chosen by design rather than derived.
  • encoder/decoder depth and width = 12/8 layers, 768/512 dims
    Le=12 blocks, De=768, Ld=8 blocks, Dd=512, 12/16 heads; standard ViT-scale choices not derived from ECG physics.
  • AdamW learning rate and schedule = 5e-4 effective
    Base 1e-3 scaled to effective 5e-4, 10-epoch warm-up, weight decay 0.05, 100 epochs, batch 128; fitted for training stability.
  • random continuous-gap mask augmentation
    Additional random temporal gaps combined with predefined clinical masks during training; strength and distribution chosen by authors.
axioms (4)
  • domain assumption Observed 12-lead samples plus a binary missingness mask supply enough temporal and inter-lead redundancy to infer missing waveform segments for typical clinical display formats.
    Stated throughout Introduction and Discussion as the physiological basis for completion; not proved, only empirically supported.
  • ad hoc to paper Simulated 4×3 2.5 s, 6×2 5 s, and local random-gap masks are sufficiently representative of real incomplete digitized ECGs for reconstruction and diagnostic-utility claims.
    Core evaluation design in Methods; Discussion acknowledges real digitization artifacts (grid, overlap, distortion) are not fully captured.
  • domain assumption L1 loss restricted to missing samples, with exact copy-back of observed samples at inference, yields clinically usable completed ECGs.
    Equations (10)–(11) in Methods; standard for imputation but assumes missing-region fidelity implies diagnostic fidelity.
  • standard math Standard Transformer self-attention and fixed sine-cosine positional embeddings are appropriate for 1D multi-lead ECG sequences.
    Architecture section; inherits ViT/MAE inductive biases without new theoretical derivation.
invented entities (1)
  • ImputeECG (mask-conditioned 1D Transformer autoencoder) independent evidence
    purpose: Unified model that conditions on observed ECG + missingness mask to reconstruct complete 12-lead 10-s waveforms while preserving observed samples exactly.
    Named architecture introduced in the paper; fully specified and code-released, so independent re-implementation is possible, but the entity itself is new to this work.

pith-pipeline@v1.1.0-grok45 · 27280 in / 3192 out tokens · 27377 ms · 2026-07-11T10:10:14.396343+00:00 · methodology

0 comments
read the original abstract

Complete digital 12-lead electrocardiograms (ECGs) are essential for AI-enabled cardiovascular assessment, yet many clinical ECG records, particularly those digitized from ECG images, remain incomplete because of short display formats, incomplete waveform digitization, lead loss, or signal corruption. We developed ImputeECG, a mask-conditioned one-dimensional Transformer autoencoder that completes 12-lead, 10-s ECGs while retaining all observed samples. The model was trained on PTB-XL and evaluated on PTB-XL and CPSC2018 under simulated incomplete settings, with additional real-world validation in a 43,633-record Kailuan clinical cohort after ECG image digitization. Metrics were computed over originally missing regions, with analyses of morphology and downstream diagnostic utility. On PTB-XL, ImputeECG reduced missing-region MAE by 41.7-51.0% and MSE by 54.0-63.7% versus the strongest baseline, with lower errors in R-peak timing, RR interval, QRS duration, QT interval, and P-wave, QRS-complex, and T-wave reconstruction. On CPSC2018, ImputeECG reduced MAE by 49.7-51.9%, supporting external generalization. In downstream multi-label classification, ImputeECG restored performance to 92.28% AUROC and 33.88% AUPRC in the most incomplete PTB-XL setting, approaching complete-ECG performance. On CPSC2018, completed ECGs achieved 94.75-95.89% AUROC and 78.83-81.86% AUPRC across settings. In Kailuan, ECG completion improved zero-shot sex prediction AUROC from 82.6% to 85.8% and reduced age prediction MAE from 10.72 to 9.87 years after image-based ECG digitization. These findings support ECG completion as a practical strategy for converting incomplete ECG records into AI-ready 12-lead, 10-s digital signals and extending the usable scope of ECG archives for digital cardiac assessment.

Figures

Figures reproduced from arXiv: 2607.05009 by Guangkun Nie, Guodong Wang, Haoyu Wang, Hongmin Liu, Hongyan Li, Jiarui Jin, Jieyi Cai, Jun Li, Liming Lin, Qinghao Zhao, Shanwei Zhang, Shenda Hong, Shouling Wu, Shun Huang, Shuohua Chen, Xiaocheng Fang, Yujie Xiao.

Figure 1
Figure 1. Figure 1: Overview of ImputeECG. a, Real-world incomplete ECG scenarios and the goal of restoring complete 12-lead, 10-second ECGs. b, Workflow of mask-guided ECG completion. c, Datasets and simulated missingness settings. d, Evaluation framework for reconstruction quality and diagnostic fidelity. e, Representative performance comparison between ImputeECG and baseline methods. including short multi-lead segments, ha… view at source ↗
Figure 2
Figure 2. Figure 2: Reconstruction performance on the PTB-XL internal test set. Bar plots compare ImputeECG with other baselines under three incomplete ECG settings. Metrics were computed over originally missing regions. Lower values indicate better performance for MAE, FD, MDD, ACD, MSE, and FID; higher values indicate better performance for SSIM and PSNR. ImputeECG showed the strongest overall reconstruction performance acr… view at source ↗
Figure 3
Figure 3. Figure 3: Morphology preservation on PTB-XL. Box plots compare reconstruction errors in ECG timing features and waveform regions across three incomplete ECG settings. Lower values indicate better preservation of ECG morphology. achieved the lowest MAE values of 0.190, 0.211, and 0.099 in the 4 × 3, 6 × 2, and 12 × 1 settings, 121 respectively, corresponding to relative reductions of 41.9%, 41.7%, and 51.0% compared … view at source ↗
Figure 4
Figure 4. Figure 4: Sex prediction performance in the independent clinical cohort. The model perfor￾mance was evaluated using completed ECGs and compared across input conditions. No Imputation Imputed by ImputeECG [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Age prediction performance in the independent clinical cohort. The model performance was evaluated using completed ECGs and compared across input conditions. setting are consistent with the fact that most lead-time information was already available and only 261 local gaps required completion. 262 The downstream analyses are central to the clinical interpretation of the study. Accurate wave- 263 form recons… view at source ↗
Figure 6
Figure 6. Figure 6: Architecture of ImputeECG. Partially observed ECGs are combined with binary missingness masks and embedded as temporal patches. A one-dimensional Transformer encoder– decoder models temporal and inter-lead dependencies to reconstruct missing regions, generating a complete 12-lead, 10-second ECG while preserving observed samples. a second-order Butterworth low-pass filter with a 50 Hz cutoff, and a 50/60 Hz… view at source ↗
Figure 7
Figure 7. Figure 7: Representative ECG completion example on PTB-XL. A 12-lead, 10-second ECG example from the 4 × 3-lead incomplete setting is shown for the masked input and reconstructions generated by EKGAN, ECGRecover, and ImputeECG. Observed waveform regions are shown in red, missing regions in the uncompleted input are shown in blue, and model-completed regions are shown in green. Compared with baseline methods, ImputeE… view at source ↗
Figure 8
Figure 8. Figure 8: Representative ECG completion example on PTB-XL in the [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Representative ECG completion example on PTB-XL in the [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 3 linked inside Pith

  1. [5]

    Performance of ImputeECG e

    Diagnostic Fidelity Multi-label Disease Classification AUROC AUPRC e. Performance of ImputeECG e. Performance of ImputeECG b. Workflow of ImputeECG Partially Missing Digital ECG Create Binary Mask ImputeECG Model Completed Digital ECG Evaluation b. Workflow of ImputeECG Partially Missing Digital ECG Create Binary Mask ImputeECG Model Completed Digital ECG...

    arXiv 2067

  2. [8]

    Comprehensive Evaluation

    Diagnostic Fidelity Multi-label Disease Classification AUROC AUPRC d. Comprehensive Evaluation

  3. [9]

    Waveform Reconstruction MAE· MSE FD· FID MDD· ACD SSIM· PSNR

  4. [10]

    Performance of ImputeECG No imputation ImputeECG e

    Diagnostic Fidelity Multi-label Disease Classification AUROC AUPRC e. Performance of ImputeECG No imputation ImputeECG e. Performance of ImputeECG No imputation ImputeECG b. Workflow of ImputeECG Partially Missing Digital ECG Create Binary Mask ImputeECG Model Completed Digital ECG Evaluation b. Workflow of ImputeECG Partially Missing Digital ECG Create B...

    2006

  5. [11]

    Kligfield, P ., Gettes, L.S., Bailey, J.J., Childers, R., Deal, B.J., Hancock, E.W., Van Herpen,507 G., Kors, J.A., Macfarlane, P ., Mirvis, D.M. et al. (2007). Recommendations for the stan- 508 dardization and interpretation of the electrocardiogram: part i: the electrocardiogram and 509 its technology: a scientific statement from the american heart asso...

    2007

  6. [12]

    Hannun, A.Y ., Rajpurkar, P ., Haghpanahi, M., Tison, G.H., Bourn, C., Turakhia, M.P ., and514 Ng, A.Y . (2019). Cardiologist-level arrhythmia detection and classification in ambulatory 515 electrocardiograms using a deep neural network. Nature medicine25, 65–69.516

    2019

  7. [13]

    Ribeiro, A.H., Ribeiro, M.H., Paixão, G.M., Oliveira, D.M., Gomes, P .R., Canazart, J.A., 517 Ferreira, M.P ., Andersson, C.R., Macfarlane, P .W., Meira Jr, W. et al. (2020). Automatic 518 diagnosis of the 12-lead ecg using a deep neural network. Nature communications11, 1760. 519

    2020

  8. [14]

    Jin, J., Fang, X., Wang, H., Li, J., Liu, C., Xie, D., Li, H., and Hong, S. (2025). Self-alignment 520 learning to improve myocardial infarction detection from single-lead ecg. arXiv preprint 521 arXiv:2509.19397. 522

    arXiv 2025

  9. [15]

    Attia, Z.I., Noseworthy, P .A., Lopez-Jimenez, F ., Asirvatham, S.J., Deshmukh, A.J., Gersh,523 B.J., Carter, R.E., Y ao, X., Rabinstein, A.A., Erickson, B.J. et al. (2019). An artificial 524 intelligence-enabled ecg algorithm for the identification of patients with atrial fibrillation 525 during sinus rhythm: a retrospective analysis of outcome predictio...

    2019

  10. [16]

    Raghunath, S., Ulloa Cerna, A.E., Jing, L., VanMaanen, D.P ., Stough, J., Hartzel, D.N., 528 Leader, J.B., Kirchner, H.L., Stumpe, M.C., Hafez, A. et al. (2020). Prediction of mortality 529 from 12-lead electrocardiogram voltage data using a deep neural network. Nature medicine 530 26, 886–891. 531

    2020

  11. [17]

    Wu, H., Patel, K.H.K., Li, X., Zhang, B., Galazis, C., Bajaj, N., Sau, A., Shi, X., Sun, L., 532 Tao, Y . et al. (2022). A fully-automated paper ecg digitisation algorithm using deep learning. 533 Scientific Reports12, 20963. 534

    2022

  12. [18]

    Jin, J., Wang, H., Wu, X., Fang, X., Lan, X., Wang, Z., Zhang, D., Liu, B., Zhang, Y ., 535 Wu, X. et al. (2026). Ecg-r1: Protocol-guided and modality-agnostic mllm for reliable ecg 536 interpretation. arXiv preprint arXiv:2602.04279. 537

    Pith/arXiv arXiv 2026

  13. [19]

    Stenhede, E., Bjørnstad, A.M., and Ranjbar, A. (2026). Digitizing paper ecgs at scale: an 538 open-source algorithm for clinical research. NPJ Digital Medicine9, 145.539

    2026

  14. [20]

    Fortune, J.D., Coppa, N.E., Haq, K.T., Patel, H., and Tereshchenko, L.G. (2022). Digitizing 540 ecg image: A new method and open-source software code. Computer methods and programs 541 in biomedicine221, 106890. 542

    2022

  15. [21]

    Gliner, V., Keidar, N., Makarov, V., Avetisyan, A.I., Schuster, A., and Y aniv, Y . (2020). 543 Automatic classification of healthy and disease conditions from images or digital standard 544 12-lead electrocardiograms. Scientific Reports10, 16331.545 21

    2020

  16. [22]

    Wagner, P ., Strodthoff, N., Bousseljot, R.D., Kreiseler, D., Lunze, F .I., Samek, W., and 546 Schaeffter, T. (2020). Ptb-xl, a large publicly available electrocardiography dataset. Scientific 547 data7, 154. 548

    2020

  17. [23]

    (2025).549 Ecgrecover: a deep learning approach for electrocardiogram signal completion

    Lence, A., Granese, F ., Fall, A., Hanczar, B., Salem, J.E., Zucker, J.D., and Prifti, E. (2025).549 Ecgrecover: a deep learning approach for electrocardiogram signal completion. In Proceed- 550 ings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1. 551 pp. 2359–2370. 552

    2025

  18. [24]

    Siontis, K.C., Noseworthy, P .A., Attia, Z.I., and Friedman, P .A. (2021). Artificial intelligence-553 enhanced electrocardiography in cardiovascular disease management. Nature Reviews 554 Cardiology18, 465–478. 555

    2021

  19. [25]

    Li, Y ., Qu, Q., Wang, M., Yu, L., Wang, J., Shen, L., and He, K. (2020). Deep learning for 556 digitizing highly noisy paper-based ecg records. Computers in biology and medicine127, 557 104077. 558

    2020

  20. [26]

    Shivashankara, K.K., Deepanshi, Mehri Shervedani, A., Clifford, G.D., Reyna, M.A., and 559 Sameni, R. (2024). Ecg-image-kit: a synthetic image generation toolbox to facilitate deep 560 learning-based electrocardiogram digitization. Physiological measurement45, 055019.561

    2024

  21. [27]

    et al.562 (2025)

    Fang, X., Jin, J., Wang, H., Liu, C., Cai, J., Xiao, Y ., Nie, G., Liu, B., Huang, S., Li, H. et al.562 (2025). Ppgflowecg: Latent rectified flow with cross-modal encoding for ppg-guided ecg 563 generation and cardiovascular disease detection. arXiv preprint arXiv:2509.19774.564

    arXiv 2025

  22. [28]

    565 et al

    Fang, X., Ding, Z., Cai, J., Xiao, Y ., Liu, B., Jin, J., Wang, H., Nie, G., Huang, S., Chen, T. 565 et al. (2026). Ecgflowcmr: Pretraining with ecg-generated cine cmr improves cardiac disease 566 classification and phenotype prediction. arXiv preprint arXiv:2601.20904.567

    Pith/arXiv arXiv 2026

  23. [29]

    Zhu, J.Y ., Park, T., Isola, P ., and Efros, A.A. (2017). Unpaired image-to-image translation 568 using cycle-consistent adversarial networks. In Proceedings of the IEEE international 569 conference on computer vision. pp. 2223–2232. 570

    2017

  24. [30]

    Isola, P ., Zhu, J.Y ., Zhou, T., and Efros, A.A. (2017). Image-to-image translation with condi-571 tional adversarial networks. In Proceedings of the IEEE conference on computer vision and 572 pattern recognition. pp. 1125–1134. 573

    2017

  25. [31]

    Golany, T., Radinsky, K., Freedman, D., and Minha, S. (2021). 12-lead ecg reconstruction 574 via koopman operators. In International Conference on Machine Learning. PMLR pp. 3745– 575

    2021

  26. [32]

    Joo, J., Joo, G., Kim, Y ., Jin, M.N., Park, J., and Im, H. (2023). Twelve-lead ecg reconstruction577 from single-lead signals using generative adversarial networks. In International Conference 578 on Medical Image Computing and Computer-Assisted Intervention. Springer pp. 184–194.579

    2023

  27. [33]

    Chen, J., Wu, W., Liu, T., and Hong, S. (2024). Multi-channel masked autoencoder and 580 comprehensive evaluations for reconstructing 12-lead ecg from arbitrary single-lead ecg. 581 NPJ Cardiovascular Health1, 34. 582

    2024

  28. [34]

    Lee, J., Y oo, Y ., Kim, J., Lim, D., Y ang, G., and Y oo, C. (2025). Parameter-efficient 12-lead ecg583 reconstruction from a single lead. In International Conference on Medical Image Computing 584 and Computer-Assisted Intervention. Springer pp. 431–441.585 22

    2025

  29. [35]

    Liu, F ., Liu, C., Zhao, L., Zhang, X., Wu, X., Xu, X., Liu, Y ., Ma, C., Wei, S., He, Z. et al. 586 (2018). An open access database for evaluating the algorithms of electrocardiogram rhythm 587 and morphology abnormality detection. Journal of Medical Imaging and Health Informatics 588 8, 1368–1373. 589

    2018

  30. [36]

    Nie, G., Tang, G., Xiao, Y ., Li, J., Huang, S., Zhang, D., Zhao, Q., and Hong, S. (2025). 590 Anyppg: An ecg-guided ppg foundation model trained on over 100,000 hours of recordings 591 for holistic health profiling. arXiv preprint arXiv:2511.01747.592

    Pith/arXiv arXiv 2025

  31. [37]

    Holmes: health online model ensemble serving for deep learning models in intensive 594 care units

    Hong, S., Xu, Y ., Khare, A., Priambada, S., Maher, K., Aljiffry, A., Sun, J., and Tumanov, A.593 (2020). Holmes: health online model ensemble serving for deep learning models in intensive 594 care units. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge 595 Discovery & Data Mining. pp. 1614–1624. 596

    2020

  32. [38]

    Koscova, Z., Li, Q., Robichaux, C., Junior, V.M., Ghanta, M., Gupta, A., Rosand, J., Aguirre, 597 A.D., Reinertsen, E., Song, S. et al. (2026). The harvard-emory ecg database. Scientific 598 Data. 599

    2026

  33. [39]

    Strodthoff, N., Wagner, P ., Schaeffter, T., and Samek, W. (2020). Deep learning for ecg 600 analysis: Benchmarks and insights from ptb-xl. IEEE journal of biomedical and health 601 informatics25, 1519–1528. 602

    2020

  34. [40]

    Li, J., Aguirre, A.D., Junior, V.M., Jin, J., Liu, C., Zhong, L., Sun, C., Clifford, G., Bran- 603 don Westover, M., and Hong, S. (2025). An electrocardiogram foundation model built on over 604 10 million recordings. Nejm ai2, AIoa2401033. 605 23

    2025

This paper was first reviewed by grok-4.5 on July 11, 2026.