Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

QualityFM: a Multimodal Physiological Signal Foundation Model with Self-Distillation for Signal Quality Challenges in Critically Ill Patients

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read QualityFM, a self-distilled ECG/PPG foundation model pretrained on 21.57 million waveform pairs, reports consistent gains over task-specific state-of-the-art models on three ICU monitoring tasks.

desk verdict A serious and well-engineered multimodal ECG/PPG foundation model whose headline 'consistently outperforms' claim is undone by its own base-model numbers and by an unresolved MIMIC pretraining/downstream overlap. read the letter →

arxiv 2509.06516 v2 pith:UQHO76QW submitted 2025-09-08 cs.LG cs.AI

Zongheng Guo , Tao Chen , Manuela Ferrario This is my paper
classification cs.LGcs.AI
keywords physiologicalsignalfoundationmodelECGPPGself-distillationqualityICUmonitoringtransferlearningsparseattention
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to show that a single pretrained model can learn a general-purpose sense of ECG and PPG signal quality and then transfer to several ICU monitoring tasks without task-specific redesign. QualityFM is pretrained on 21.57 million paired 30-second segments from two public waveform databases, using pairs of high- and low-quality signals recorded close in time from the same patient. A teacher encoder sees the high-quality segment and distills its representation into a student encoder that sees the low-quality one, while a shared decoder reconstructs the amplitude and phase spectra of the signal. After fine-tuning, the same pretrained encoder improves false ventricular tachycardia alarm detection, atrial fibrillation identification, and arterial blood pressure estimation compared with state-of-the-art task-specific models. The claim matters because it points toward a label-efficient, quality-aware backbone for bedside monitoring.

What carries the argument

The load-bearing mechanism is a dual-track self-distillation loop over quality-divergent pairs. For each pair, the high-quality segment feeds a teacher encoder updated by exponential moving average; the low-quality segment feeds a student encoder trained to match the teacher's soft output distribution. A decoder reconstructs the high-quality segment's amplitude and phase spectra from the student features, forcing the student to preserve frequency-domain morphology. A windowed sparse attention Transformer with window size 8 keeps the 9000-sample inputs tractable while capturing local waveform shape. Together these components define the pretraining objective that the authors claim creates tran

What would settle it

Compute the exact patient-level overlap between the pretraining corpora (VitalDB and the MIMIC-III Matched Subset) and the three downstream evaluation sets (VTaC, MIMIC PERform AF, UCI BP), then fine-tune on a strictly patient-disjoint split. If the state-of-the-art-level gains disappear or shrink substantially under that split, the central claim of generalizable transfer is not supported.

Watch

Extended reading notes

Core claim

The central claim is that pretraining on signal-quality-divergent pairs yields representations that transfer better than existing quality-focused contrastive learning and task-specific architectures. The model family scales from 9.6M to 319M parameters; the largest variant consistently improves all metrics on the VTaC false-alarm task (accuracy 0.8551 vs 0.7432 for the best reported baseline), on MIMIC PERform AF (accuracy 0.8764 vs 0.8457), and lowers blood-pressure estimation MAE on the UCI dataset (SBP 19.32 vs 20.64; DBP 7.61 vs 8.32). The authors interpret this as evidence that a scalable, general-purpose quality-aware representation is learnable from unlabeled waveforms and is transfer

Load-bearing premise

The central claim stands on the assumption that the patients and waveform segments used for pretraining are disjoint from those used for downstream evaluation; if the same patients appear in both, the reported transfer gains are inflated.

Editorial extensions

If this is right

  • If the claim holds, downstream ICU tasks can be built by fine-tuning a single pretrained ECG/PPG encoder rather than training task-specific networks from scratch.
  • The consistent gains from 9.6M to 319M parameters suggest further scaling of pretraining data and model size would continue to improve transfer, assuming the data-disjointness condition is met.
  • The pretraining objective requires no expert labels: quality pseudo-labels are computed from signal-quality indices, so the approach can be replicated on any paired waveform archive.
  • Frequency-domain reconstruction makes the learned features sensitive to waveform morphology, which is directly relevant to alarm classification and blood-pressure estimation.
  • A quality-aware backbone could reduce false alarms and alarm fatigue in ICU monitoring if deployed with the same preprocessing pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The decisive untested check is patient-level separation between pretraining and downstream data; the authors do not report whether the MIMIC-III pretraining subset and the MIMIC-derived VTaC, PERform AF, and UCI test sets share patients. If they do, the transfer gains would shrink when evaluated on unseen patients.
  • The quality-divergent pairing idea is not limited to ECG and PPG: any quasi-periodic biosignal with a computable quality index, such as respiratory effort or fetal heart rate, could be pretrained in the same teacher-student loop.
  • The windowed sparse attention results suggest that choosing the right receptive field matters for local waveform morphology, so smaller variants may suffice in resource-constrained bedside monitors.
  • A natural extension the authors leave implicit is evaluating QualityFM under distribution shift to different monitor hardware or sampling rates; the fixed 300 Hz resampling leaves robustness to hardware differences an open question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces QualityFM, a multimodal ECG/PPG foundation model pretrained via self-distillation on quality-divergent segment pairs from VitalDB and the MIMIC-III Waveform Database Matched Subset. A teacher encoder processes high-quality segments and guides a student encoder on low-quality segments via an EMA-updated distillation objective, with a windowed sparse attention transformer and a composite loss combining distillation with amplitude/phase spectral reconstruction. Three model sizes (9.6M, 70M, 319M parameters) are pretrained on 21.57 million 30-second segments and then fine-tuned on three downstream tasks: ventricular tachycardia false-alarm detection (VTaC), atrial fibrillation identification (MIMIC PERform AF), and arterial blood pressure estimation (UCI). The paper claims consistent state-of-the-art performance and improved transfer from pretraining, with ablations supporting the windowed attention and spectral reconstruction components.

Significance. If the reported results are valid, this would be a substantial contribution: a large-scale, signal-quality-aware foundation model for ECG/PPG with a novel self-distillation formulation and credible transfer to multiple ICU monitoring tasks. The scale of pretraining (179,757 hours), the public benchmarks used, and the systematic ablations are strengths. However, the current evidence is weakened by an unresolved potential overlap between pretraining and downstream data, the absence of uncertainty quantification in all comparative tables, and internal inconsistencies in the claimed 'consistent outperformance' and scaling trends. These issues are central to the paper's main claims and must be addressed before the results can be accepted.

major comments (4)
  1. [Datasets for Pretraining Task / Results - Training and Evaluation] The pretraining corpus includes the MIMIC-III Waveform Database Matched Subset (10,282 patients), while all three downstream datasets (VTaC, MIMIC PERform AF, and UCI BP) are MIMIC-derived. The manuscript never states that pretraining and downstream subjects are disjoint at the patient level, nor that the 80/20 splits for MIMIC PERform AF and UCI are patient-stratified. Because pretraining uses overlapping 30-second windows from the same 125/300 Hz streams, any patient overlap could allow the model to memorize subject-specific morphology and inflate the transfer gains in Tables 1 and 2. This is load-bearing for every comparative claim. A patient-level overlap audit and, if needed, retraining with exclusion or stratified splits are required.
  2. [Tables 1 and 2] All performance numbers are reported as point estimates without error bars, confidence intervals, or significance tests. For MIMIC PERform AF and UCI, the text says an 80/20 split with 'cross-validation methodology' was used, but no repeated-seed variance, patient-stratification details, or per-fold results are given. Without this information, the statements 'consistent and significant improvement' and 'consistently outperforms' are not statistically supported. Reporting mean±std over multiple seeds and patient-stratified splits is necessary.
  3. [Table 1, VTaC rows] The text in 'Comparison with the State-of-the-Art performances' states that 'QualityFM-Base consistently outperforms the existing SOTA approaches on both datasets.' This is contradicted by Table 1: on VTaC, QualityFM-Base AUC is 0.8565, below ResNet-18's 0.9097. Additionally, QualityFM-Huge AUC is 0.9047, lower than QualityFM-Large AUC 0.9099, which contradicts the claim of 'consistent and significant improvement' with increasing model size. The claims need to be revised to match the tabulated numbers, and the scaling trend needs a more careful description.
  4. [Table 2] The last two rows of Table 2 are both labeled 'QualityFM-Large' with different values (SBP MAE 19.46 vs 19.32; DBP MAE 7.86 vs 7.61). This appears to be a typo, likely the second row should be 'QualityFM-Huge'. The duplication prevents verification of the scaling trend and of the claim that larger models improve accuracy. The table must be corrected.
minor comments (4)
  1. [Throughout] There are many typographical errors, including 'intesive', 'updatpte', 'udapte', 'metrci', 'teh', 'segmnet', and 'taks'. The paper needs a careful proofreading pass.
  2. [Appendix D, Table 5] Table 5 lists 'Momentum 0.7', but Eq. (4) in the main text says the EMA momentum parameter λ follows a cosine schedule increasing from 0.996 to 1.0. Please clarify which 'momentum' is meant and reconcile the discrepancy.
  3. [Results, Model naming] The text uses 'PSW A-Base', 'PSW A-Large', and 'PSW A-Huge' in the Training and Evaluation paragraph, while the tables use 'QualityFM-Base/Large/Huge'. Define the acronym PWSA and use the naming consistently.
  4. [Eq. (5)] In the attention formula, the pre-LN formulation is written with LN(Q) and LN(K), but it is unclear whether LN is applied per-head or per-token, and whether the scaling by √d_head is applied before or after LN. A precise definition would aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: downstream labels are independent of the pretraining objective; spectral reconstruction targets are derived from input signals, and self-citations are not load-bearing.

full rationale

The central derivation chain is self-contained. The pretraining loss (Eq. 9) combines a self-distillation term over quality-divergent pairs with spectral-reconstruction losses whose targets are DFT amplitude and phase spectra computed from the input signals (Eqs. 6-8) — not from any downstream label. Downstream tasks use external labels: VTaC false-alarm labels, MIMIC PERform AF rhythm labels, and UCI ABP values, none of which appear in the pretraining objective. The hand-crafted SQI thresholds (Eq. 10) only define the pairing criterion; they do not define the downstream targets. The from-scratch ablation provides independent evidence for the pretraining benefit. There are self-citations (e.g., Guo et al. 2023; Ding et al. 2024), but they are used for context or as baselines and are not load-bearing for the main transfer claims. The patient-overlap/leakage concern between MIMIC-based pretraining and MIMIC-derived downstream datasets is a validity threat, not a circular reduction; no equation or construction forces the downstream predictions to equal the pretraining objective. The statement that QualityFM-Base 'consistently outperforms' SOTA is factually contradicted by Table 1 (QualityFM-Base VTaC AUC 0.8565 < ResNet-18 0.9097), but this is a correctness issue, not circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The model's empirical performance rests entirely on hand-crafted signal-quality heuristics (SQI) and on the assumption that public MIMIC waveform data can be used for both pretraining and downstream evaluation without leakage. No new physical entities are introduced.

free parameters (5)
  • window_size = 8
    Sparse attention window size selected via downstream-task ablation on all three datasets (Figure 3). Affects all reported results.
  • lambda_Amp = 0.5
    Weight for amplitude reconstruction loss in composite loss (Appendix D, Table 5).
  • lambda_Pha = 0.5
    Weight for phase reconstruction loss in composite loss (Appendix D, Table 5).
  • SQI_quality_thresholds = [0.9, 0.7, 0.5, 0.3]
    Categorical boundaries for quality pseudo-labels in Equation (10); hand-set and used to define pretraining pairs.
  • ema_momentum_start = 0.996 (text); Table 5 lists Momentum 0.7
    EMA momentum schedule for teacher updates; textual description and hyperparameter table are inconsistent.
assumptions (4)
  • domain assumption The computed Signal Quality Index (SQI) heuristics correctly reflect true signal quality.
    Used to define quality labels in Equation (10) and to select quality-divergent pretraining pairs; if SQI is wrong, the self-distillation objective is misaligned. (Appendix C)
  • domain assumption High-quality segments have higher information density and are reliable teachers for low-quality segments.
    Central to the self-distillation design in Methodology.
  • domain assumption Paired segments from the same subject within 5 minutes with different quality labels provide useful training signal.
    Pair selection criterion in Equation (1).
  • domain assumption The pretraining data (VitalDB, MIMIC-III) and downstream data (VTaC, MIMIC PERform AF, UCI) have no subject overlap that would inflate transfer performance.
    The paper never states whether subject-level de-contamination was performed; if overlap exists, downstream gains are overstated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QualityFM: a Multimodal Physiological Signal Foundation Model with Self-Distillation for Signal Quality Challenges in Critically Ill Patients." pith.science (2026). https://pith.science/paper/UQHO76QW

@misc{pith2026250906516,
  author       = {Pith},
  title        = {Pith review of: QualityFM: a Multimodal Physiological Signal Foundation Model with Self-Distillation for Signal Quality Challenges in Critically Ill Patients},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQHO76QW}},
  note         = {Machine review of arXiv:2509.06516}
}
read the original abstract

Photoplethysmogram (PPG) and electrocardiogram (ECG) are commonly recorded in intesive care unit (ICU) and operating room (OR). However, the high incidence of poor, incomplete, and inconsistent signal quality, can lead to false alarms or diagnostic inaccuracies. The methods explored so far suffer from limited generalizability, reliance on extensive labeled data, and poor cross-task transferability. To overcome these challenges, we introduce QualityFM, a novel multimodal foundation model for these physiological signals, designed to acquire a general-purpose understanding of signal quality. Our model is pre-trained on an large-scale dataset comprising over 21 million 30-second waveforms and 179,757 hours of data. Our approach involves a dual-track architecture that processes paired physiological signals of differing quality, leveraging a self-distillation strategy where an encoder for high-quality signals is used to guide the training of an encoder for low-quality signals. To efficiently handle long sequential signals and capture essential local quasi-periodic patterns, we integrate a windowed sparse attention mechanism within our Transformer-based model. Furthermore, a composite loss function, which combines direct distillation loss on encoder outputs with indirect reconstruction loss based on power and phase spectra, ensures the preservation of frequency-domain characteristics of the signals. We pre-train three models with varying parameter counts (9.6 M to 319 M) and demonstrate their efficacy and practical value through transfer learning on three distinct clinical tasks: false alarm of ventricular tachycardia detection, the identification of atrial fibrillation and the estimation of arterial blood pressure (ABP) from PPG and ECG signals.

Figures

Figures reproduced from arXiv: 2509.06516 by the authors.

Figure 1
Figure 1. Pipeline of QualityFM pretraining and downstream tasks. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Comparison of the model performances with and [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Comparison of QualityFM with different back [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Example of ECG and PPG signals with different [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. When Derived Measurements Mislead: Quantifying and Mitigating LLM Over-Trust with Privileged-Modality Reliability Evidence

    cs.AI 2026-07 conditional novelty 6.5 of 10

    DFOT is formalized with five frozen estimands and a matched–shuffled evidence test; on locked PPG–ECG data a privileged student modestly improves repair and specificity over a PPG-only baseline.

Reference graph

Works this paper leans on

46 extracted references · 33 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    J.; and Armentano, R

    Aguirre, N.; Grall-Ma \"e s, E.; Cymberknop, L. J.; and Armentano, R. L. 2021. Blood pressure morphology assessment from photoplethysmogram and demographic information using deep learning with attention mechanism. Sensors, 21(6): 2167

    work page 2021

  4. [4]

    A.; Islam, M

    Almarshad, M. A.; Islam, M. S.; Al-Ahmadi, S.; and BaHammam, A. S. 2022. Diagnostic features and potential applications of PPG signal in healthcare: A systematic review. In Healthcare, volume 10, 547. MDPI

    work page 2022

  5. [5]

    Balaji, S.; Ellenby, M.; McNames, J.; and Goldstein, B. 2002. Update on intensive care ECG and cardiac event monitoring. Cardiac electrophysiology review, 6: 190--195

    work page 2002

  6. [6]

    A.; and Nayak, R

    Bashar, M. A.; and Nayak, R. 2020. TAnoGAN: Time series anomaly detection with generative adversarial networks. In 2020 IEEE Symposium Series on Computational Intelligence (SSCI), 1778--1785. IEEE

    work page 2020

  7. [7]

    K.; Ding, E.; Walkey, A

    Bashar, S. K.; Ding, E.; Walkey, A. J.; McManus, D. D.; and Chon, K. H. 2019. Noise detection in electrocardiogram signals for intensive care unit patients. IEEE Access, 7: 88357--88368

    work page 2019

  8. [8]

    Behar, J.; Oster, J.; Li, Q.; and Clifford, G. D. 2013. ECG signal quality during arrhythmia and its application to false alarm reduction. IEEE transactions on biomedical engineering, 60(6): 1660--1666

    work page 2013

Show all 46 references
  1. [9]

    E.; and Cohan, A

    Beltagy, I.; Peters, M. E.; and Cohan, A. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150

    2020 arXiv

  2. [10]

    H.; Borges, R

    Borghi, P. H.; Borges, R. C.; and Teixeira, J. P. 2021. Atrial fibrillation classification based on MLP networks by extracting Jitter and Shimmer parameters. Procedia Computer Science, 181: 931--939

    2021

  3. [11]

    Caron, M.; Touvron, H.; Misra, I.; J \'e gou, H.; Mairal, J.; Bojanowski, P.; and Joulin, A. 2021. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 9650--9660

    2021

  4. [12]

    Chen, H.; Zhang, X.; Guo, Z.; Ying, N.; Yang, M.; and Guo, C. 2024. ACTNet: attention based CNN and Transformer network for respiratory rate estimation. Biomedical Signal Processing and Control, 96: 106497

    2024

  5. [13]

    Chen, Z.; Ding, C.; Kataria, S.; Yan, R.; Wang, M.; Lee, R.; and Hu, X. 2025. GPT-PPG: a GPT-based foundation model for photoplethysmography signals. Physiological Measurement, 46(5): 055004

    2025

  6. [14]

    P.; Caron, M.; Geirhos, R.; Alabdulmohsin, I.; et al

    Dehghani, M.; Djolonga, J.; Mustafa, B.; Padlewski, P.; Heek, J.; Gilmer, J.; Steiner, A. P.; Caron, M.; Geirhos, R.; Alabdulmohsin, I.; et al. 2023. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, 7480--7512. PMLR

    2023

  7. [15]

    J.; Rudin, C.; and Hu, X

    Ding, C.; Guo, Z.; Chen, Z.; Lee, R. J.; Rudin, C.; and Hu, X. 2024. SiamQuality: a ConvNet-based foundation model for photoplethysmography signals. Physiological Measurement, 45(8): 085004

    2024

  8. [16]

    J.; Harris, P.; Z \`e gre-Hemsey, J

    Drew, B. J.; Harris, P.; Z \`e gre-Hemsey, J. K.; Mammone, T.; Schindler, D.; Salas-Boni, R.; Bai, Y.; Tinoco, A.; Ding, Q.; and Hu, X. 2014. Insights into the problem of alarm fatigue with physiologic monitor devices: a comprehensive observational study of consecutive intensi...

    2014

  9. [17]

    Gambarotta, N.; Aletti, F.; Baselli, G.; and Ferrario, M. 2016. A review of methods for the signal quality assessment to improve reliability of heart rate and blood pressures derived parameters. Medical & biological engineering & computing, 54: 1025--1035

    2016

  10. [18]

    Gonz \'a lez, S.; Hsieh, W.-T.; and Chen, T. P.-C. 2023. A benchmark for machine-learning based non-invasive blood pressure estimation using photoplethysmogram. Scientific Data, 10(1): 149

    2023

  11. [19]

    Guo, Z.; Chen, H.; Lin, L.; Zhou, W.; Yang, M.; Ying, N.; and Guo, C. 2023. Remote heart rate estimation via convolutional neural networks with transformers. Journal of the Franklin Institute, 360(17): 13149--13165

    2023

  12. [20]

    He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770--778

    2016

  13. [21]

    L.; Rakocz, N.; Rudas, \'A .; Chiang, J

    Hill, B. L.; Rakocz, N.; Rudas, \'A .; Chiang, J. N.; Wang, S.; Hofer, I.; Cannesson, M.; and Halperin, E. 2021. Imputation of the continuous arterial line blood pressure waveform from non-invasive measurements using deep learning. Scientific reports, 11(1): 15755

    2021

  14. [22]

    Huang, Z.; Xu, W.; and Yu, K. 2015. Bidirectional LSTM-CRF models for sequence tagging. arXiv preprint arXiv:1508.01991

    2015 arXiv

  15. [23]

    E.; Pollard, T

    Johnson, A. E.; Pollard, T. J.; Shen, L.; Lehman, L.-w. H.; Feng, M.; Ghassemi, M.; Moody, B.; Szolovits, P.; Anthony Celi, L.; and Mark, R. G. 2016. MIMIC-III, a freely accessible critical care database. Scientific data, 3(1): 1--9

    2016

  16. [24]

    M.; Mohammadzade, H.; and Shabany, M

    Kachuee, M.; Kiani, M. M.; Mohammadzade, H.; and Shabany, M. 2015. Cuff-less high-accuracy calibration-free blood pressure estimation using pulse transit time. In 2015 IEEE international symposium on circuits and systems (ISCAS), 1006--1009. IEEE

    2015

  17. [25]

    B.; Yang, S

    Lee, H.-C.; Park, Y.; Yoon, S. B.; Yang, S. M.; Park, D.; and Jung, C.-W. 2022. VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients. Scientific Data, 9(1): 279

    2022

  18. [26]

    Lehman, L.-w.; Moody, B.; Deep, H.; Wu, F.; Saeed, H.; McCullum, L.; Perry, D.; Struja, T.; Li, Q.; Clifford, G.; et al. 2023. VTaC: A benchmark dataset of ventricular tachycardia alarms from ICU monitors. Advances in Neural Information Processing Systems, 36: 38827--38843

    2023

  19. [27]

    Li, Q.; and Clifford, G. D. 2012. Signal quality and data fusion for false alarm reduction in the intensive care unit. Journal of electrocardiology, 45(6): 596--603

    2012

  20. [28]

    Liu, G.; Han, X.; Tian, L.; Zhou, W.; and Liu, H. 2021. ECG quality assessment based on hand-crafted statistics and deep-learned S-transform spectrogram features. Computer Methods and Programs in Biomedicine, 208: 106269

    2021

  21. [29]

    Liu, L.; Cui, G.; Wan, C.; Wu, D.; and Li, Y. 2024. ECG-LLM: Leveraging Large Language Models for Low-Quality ECG Signal Restoration. In 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 3537--3542. IEEE

    2024

  22. [30]

    McKeen, K.; Oliva, L.; Masood, S.; Toma, A.; Rubin, B.; and Wang, B. 2024. Ecg-fm: An open electrocardiogram foundation model. arXiv preprint arXiv:2408.05178

    2024 arXiv

  23. [31]

    Mousavi, S.; Fotoohinasab, A.; and Afghah, F. 2020. Single-modal and multi-modal false arrhythmia alarm reduction using attention-based convolutional and recurrent neural networks. PloS one, 15(1): e0226990

    2020

  24. [32]

    Oster, J.; Behar, J.; Colloca, R.; Li, Q.; Li, Q.; and Clifford, G. D. 2013. Open source Java-based ECG analysis software and Android app for atrial fibrillation screening. In Computing in Cardiology 2013, 731--734. IEEE

    2013

  25. [33]

    Rim, B.; Sung, N.-J.; Min, S.; and Hong, M. 2020. Deep learning in physiological signal data: A survey. Sensors, 20(4): 969

    2020

  26. [34]

    Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention--MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part II...

    2015

  27. [35]

    Slapni c ar, G.; Mlakar, N.; and Lu s trek, M. 2019. Blood pressure estimation from photoplethysmogram using a spectro-temporal deep neural network. Sensors, 19(15): 3420

    2019

  28. [36]

    Tang, W.; Long, G.; Liu, L.; Zhou, T.; Jiang, J.; and Blumenstein, M. 2020. Rethinking 1d-cnn for time series classification: A stronger baseline. arXiv preprint arXiv:2002.10061, 1--7

    2020 arXiv

  29. [37]

    van der Bijl, K.; Elgendi, M.; and Menon, C. 2022. Automatic ECG quality assessment techniques: A systematic review. Diagnostics, 12(11): 2578

    2022

  30. [38]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  31. [39]

    Wan, Z.; Yu, Q.; Mao, J.; Duan, W.; and Ding, C. 2025. Openecg: Benchmarking ecg foundation models with public 1.2 million records. arXiv preprint arXiv:2503.00711

    2025 arXiv

  32. [40]

    Wu, F.; Zhao, G.; Qian, X.; and Lehman, L.-W. H. 2023. A Diffusion Model with Contrastive Learning for ICU False Arrhythmia Alarm Reduction. In IJCAI, 4912--4920

    2023

  33. [41]

    J.; Nahab, F.; and Hu, X

    Yan, R.; Ding, C.; Xiao, R.; Fedorov, A.; Lee, R. J.; Nahab, F.; and Hu, X. 2024. SQUWA: signal quality aware DNN architecture for enhanced accuracy in atrial fibrillation detection from noisy PPG signals. arXiv preprint arXiv:2404.15353

    2024 arXiv

  34. [42]

    Zhang, H.; Liu, W.; Shi, J.; Chang, S.; Wang, H.; He, J.; and Huang, Q. 2022. Maefe: Masked autoencoders family of electrocardiogram for self-supervised pretraining and transfer learning. IEEE Transactions on Instrumentation and Measurement, 72: 1--15

    2022

  35. [43]

    Zhou, B.; Liu, S.; Hooi, B.; Cheng, X.; and Ye, J. 2019. Beatgan: Anomalous rhythm detection using adversarially generated time series. In IJCAI, volume 2019, 4433--4439

    2019

  36. [44]

    Zhou, J.; Wei, C.; Wang, H.; Shen, W.; Xie, C.; Yuille, A.; and Kong, T. 2021. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832

    2021 arXiv

  37. [45]

    G.; and Lehman, L.-w

    Zhou, Y.; Zhao, G.; Li, J.; Sun, G.; Qian, X.; Moody, B.; Mark, R. G.; and Lehman, L.-w. H. 2022. A contrastive learning approach for ICU false arrhythmia alarm reduction. Scientific reports, 12(1): 4689

    2022

  38. [46]

    Zong, W.; Moody, G.; and Jiang, D. 2003. A robust open-source algorithm to detect onset and duration of QRS complexes. In Computers in Cardiology, 2003, 737--740. IEEE

    2003

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.