Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

FoundationalECGNet: A Lightweight Foundational Model for ECG-based Multitask Cardiac Analysis

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

Pith's one-line read FoundationalECGNet claims a single lightweight model can screen ECG signals, then assign one of five cardiac conditions, at about 99% F1.

desk verdict The headline numbers don't survive contact with the paper's own protocol — the ADASYN-before-split step alone is enough to make the reported F1s untrustworthy. read the letter →

arxiv 2509.08961 v1 pith:MMYV3SQ4 submitted 2025-09-10 cs.LG

classification cs.LG
keywords ECGclassificationfoundationalmodelwaveletdenoisinggraphattentionnetworktimeseriestransformermultitaskcardiacanalysisclassimbalanceneuro-symbolic
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 proposes FoundationalECGNet, a compact deep-learning model that first screens ECG recordings as Normal or Abnormal and, if Abnormal, assigns one of five cardiac conditions: arrhythmia, conduction disorder, myocardial infarction, QT abnormality, or hypertrophy. The model uses two wavelet transforms to clean the signal, then a fusion of convolutional, graph-attention, and transformer blocks to capture inter-lead and long-range temporal patterns. Across four public datasets, the paper reports 99% F1 for Normal/Abnormal separation and 98.9–99% F1 for several disease classes, with about 7.5 million parameters. The authors argue this makes automated ECG interpretation accurate, interpretable, and light enough for wearable or resource-limited use.

What carries the argument

The load-bearing mechanism is late fusion of four complementary feature extractors: Morlet and DB4 wavelets plus ExBlock convolutions produce a denoised latent set; CBAM reweights channel and spatial features; GAT treats the ECG leads as nodes of a graph and learns inter-lead attention; TST applies multi-head self-attention over time. Their outputs are concatenated, pooled, and classified, and a rule-based symbolic module turns detected intervals into disease probabilities. This combination is what the paper credits for both accuracy and interpretability.

What would settle it

Split each dataset first, then run ADASYN only on the training folds, retrain FoundationalECGNet, and compare F1 on the untouched test fold. If the scores stay at 99%, the accuracy claim is robust; if they drop substantially on minority classes, the leakage explains the headline numbers.

Watch

Extended reading notes

Core claim

The paper's central discovery is an architecture that combines dual-stage wavelet denoising (Morlet and Daubechies DB4) with an ExBlock convolutional stem, CBAM spatial-channel attention, a Graph Attention Network over ECG leads, and a Time Series Transformer, fusing their outputs into a shared latent space before classification. This mixture is claimed to preserve both local waveform details and long-range temporal context, which the authors say is why the model reaches near-ceiling F1 scores on difficult classes such as myocardial infarction and QT abnormalities while staying at 7.5 million parameters. A symbolic rule layer then maps detected P-Q-R-S-T features and intervals into disease p

Load-bearing premise

The reported numbers assume that ADASYN oversampling is confined to the training set, but the paper's pseudo-code applies it to the pooled features before any train/test split, so synthetic copies of test cases can enter training and inflate every F1 score.

Editorial extensions

If this is right

  • If the reported numbers hold, one lightweight model can serve as both a screening gate and a triage tool: it first separates normal from abnormal, then identifies which of five high-risk conditions is present.
  • Because the model is pre-trained sequentially on several public datasets and then fine-tuned on a target set, its success would suggest that cross-dataset ECG training is viable without a private corpus.
  • The symbolic scoring layer means the model outputs not just a label but a confidence and a rule-based rationale, which could make AI-assisted ECG review more auditable in clinical settings.
  • At 7.5 million parameters and 28.62 MB, the architecture is small enough that the same approach could be packaged for on-device, real-time monitoring rather than cloud inference.
  • The ablation results identify TST and GAT as the largest contributors to performance, implying that long-range temporal and inter-lead relationships carry most of the diagnostic signal.

Reading between the lines

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

  • A natural next test is to replace the hand-authored symbolic rules with a learned consistency constraint, forcing the network and the rule module to agree; this would reveal whether the symbolic layer adds signal or merely echoes the neural output.
  • If the architecture's success transfers, the same graph-of-channels plus transformer-over-time recipe could be applied to other multi-channel biosignals such as EEG, where lead/channel relationships and long-range dynamics are also central.
  • The reported near-100% figures should be re-checked under a strictly leakage-free protocol, applying ADASYN only within the training fold; this is the single most direct way to see whether the balancing step, rather than the architecture, is responsible for the high F1 scores.
  • One could benchmark the two-stage wavelet denoiser against a learned denoiser (e.g., an autoencoder) with the same downstream classifier to quantify how much of the gain comes from the classical pre-processing step.
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 proposes FoundationalECGNet, a multi-component neural architecture combining dual-stage wavelet denoising (Morlet and Daubechies DB4), ExBlock convolutional blocks, CBAM attention, Graph Attention Networks, and a Time Series Transformer for hierarchical ECG classification: first Normal vs. Abnormal, then one of five disease classes (Arrhythmia, Conduction Disorder, MI, QT Abnormality, Hypertrophy). A rule-based 'neuro-symbolic' module computes risk scores from hand-crafted thresholds. The authors report near-perfect F1 scores on a PTB test set, claim state-of-the-art performance, and emphasize the model's lightweight footprint (7.5M parameters). The architecture is described in algorithmic detail, but the evaluation protocol contains a critical data-leakage flaw and the reported metrics are internally inconsistent.

Significance. If the reported results were valid, FoundationalECGNet would be a promising lightweight multitask ECG classifier with a rare combination of high accuracy and interpretable risk outputs. The architectural components are individually well-known, and the computational-efficiency comparison against larger foundation models is useful. However, the evidence presented for the central performance claim is invalidated by a pre-split ADASYN application and by unexplainable macro-averaged scores. The paper's value is therefore contingent on a complete re-evaluation; in its current state it does not substantiate the claimed state-of-the-art performance.

major comments (4)
  1. [Algorithm 1, lines 20-23] ADASYN is applied to the full feature matrix C and labels Y before any train/test split, and the model is trained (line 22) before the test split is created (line 23). This is a direct data-leakage path: synthetic samples generated from pooled data can encode information from examples that later appear in X_test, inflating all reported metrics. The ordering is also internally inconsistent—training cannot precede test-set construction. The reported F1 scores in Tables II and III cannot be interpreted as generalization estimates. A valid protocol must fix the test split and apply ADASYN only to the training fold.
  2. [Table II] The Macro-Avg row for FoundationalECGNet (Precision=1.000, Recall=0.990, F1=0.995) contradicts the per-class values in the same table. Averaging the five per-class F1 values (0.989, 0.990, 0.927, 0.939, 0.990) yields 0.967, not 0.995; similar discrepancies appear for precision and recall. This arithmetic inconsistency is not a minor typo—it indicates that the aggregate metrics were computed from a different set of predictions or by an erroneous formula, further undermining the reported results.
  3. [Section III-B, Eq. (15)-(19)] The disease 'probabilities' and risk levels in Eqs. (15)-(19) are fixed hand-crafted functions of hard thresholds (e.g., |HR−75|/50, PR>200 ms, QRS>120 ms, ST>0.1 mV, QT>460 or <350 ms, R>2.0 mV) combined through a priority ordering. No calibration, uncertainty quantification, or clinical validation is provided, so calling these outputs 'probabilities' is unjustified. The paper's interpretability and risk-assessment claims rest on this module, yet the manuscript offers no evidence that these scores are accurate or clinically meaningful. At minimum, they should be labeled heuristic scores and validated against held-out expert labels or outcomes.
  4. [Section IV and Algorithm 1] The test-set construction is ambiguous and appears contradictory. Section IV states the first three datasets are split into train/validation and the final dataset (PTB) is split into train/test, suggesting a dataset-wise split. Algorithm 1, however, pools all data (C, Y) and applies ADASYN before any split, implying a random subject-level split that would mix datasets. This inconsistency makes it impossible to determine what the test set actually contains, which is essential for interpreting the claimed 99.48% accuracy.
minor comments (4)
  1. [General] The manuscript states 'We'll share the code after acceptance.' Releasing code and the exact data-preprocessing pipeline would greatly aid reproducibility, especially given the ambiguity in the splitting procedure.
  2. [Figure 4] The '3D confidence level' and the annotated waveform analysis are described only briefly; a clearer explanation of how the visualizations are generated and how clinicians should interpret them is needed.
  3. [Table III] The ablation table reports inconsistent decimal places (e.g., 0.9948 vs. 0.9948, but 0.9372 vs. 0.9370). Standardize the precision and ensure each row's values are derived from the same number of significant digits.
  4. [Eq. (4)] Equation (4) defines F_DB4 = f_high(t) · f_low(t) without justification. Element-wise multiplication of approximation and detail coefficients is not a standard wavelet reconstruction; please explain why this fusion is valid or replace it with a standard reconstruction method (e.g., inverse DWT).

Circularity Check

1 steps flagged · score 6.0 of 10

Pre-split ADASYN contaminates the test set, making reported F1 scores an artifact of training on synthetic test-neighbor data.

  1. fitted input called prediction [Algorithm 1, lines 20-23 (Section III-A / IV)]
    "Y ← ADASYN(C, Y); Xtrain, Xval, Ytrain, Yval ← trainValSplit(C, Y); Train model M ← FoundationalECGNet(); Xtrain, Xtest, Ytrain, Ytest ← trainTestSplit(C, Y, test_size=0.2)"

    ADASYN is applied to the combined feature matrix C and labels Y before any split. ADASYN synthesizes samples based on k-nearest neighbors in the full dataset, so synthetic training samples are generated from a distribution that includes test instances. The subsequent trainTestSplit then creates a training set containing synthetic near-duplicates of test samples, meaning the model is trained on test-like data before 'predicting' the test set. The reported F1 scores (e.g., 0.989-0.990 in Table II) are therefore inflated by construction and do not measure generalization; the 'prediction' on the test set is statistically forced by the training protocol.

full rationale

The paper's central claim of state-of-the-art ECG classification rests on the evaluation protocol in Algorithm 1. Lines 20-23 apply ADASYN to the combined feature matrix and labels before any split, then perform the train/test split from the augmented data. Because ADASYN synthesizes samples around nearest neighbors in the full dataset, the training set can contain near-duplicates of later test instances, creating a direct leakage path. The reported test metrics are therefore not independent evidence for the model's generalization; they reduce, by construction, to the pre-split augmentation rather than to learned generalizable representations. No other significant circularity was found: the wavelet transforms, attention modules, and symbolic thresholds are either standard formulas, design choices, or external clinical criteria. The only self-citation [47] is for the standard attention equation and is not load-bearing. The main issue is the data leakage, which warrants a score of 6.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The model's output depends on numerous hand-picked thresholds and heuristic formulas for the rule-based risk scores, on an unsafe oversampling protocol, and on several unvalidated signal processing assumptions. The neural network's architecture is a composition of known modules.

free parameters (9)
  • R-peak detection threshold = 0.3 * max(s)
    Eq. 14 uses this ratio to detect R peaks; it is a hand-set threshold.
  • Minimum R-peak distance = 0.4 * fs
    Eq. 14 imposes a minimum separation; hand-set.
  • Arrhythmia heart-rate offset and scale = 75 bpm and 50 bpm
    Eq. 15 computes P_AM from |HR-75|/50; these constants are arbitrary and not derived.
  • Conduction disorder PR and QRS thresholds = 200 ms, 120 ms
    Eq. 16 uses clinical thresholds as fixed parameters.
  • MI ST elevation and QRS thresholds = 0.1 mV, 100 ms
    Eq. 17 uses these thresholds with additional scale factors.
  • QT interval boundaries = 460 ms, 350 ms
    Eq. 18 uses these limits as fixed cutoffs.
  • Hypertrophy R amplitude and QRS thresholds = 2.0 mV, 110 ms
    Eq. 19 uses these thresholds and scaling constants.
  • Disease priority order = MI > AM > CD > QT > HE
    Section III states priority-based selection; this ordering is a design choice not derived.
  • Morlet frequency factor = 5
    Eq. 1 uses cos(5t) chosen to match QRS frequencies; hand-set.
assumptions (4)
  • ad hoc to paper ADASYN applied to the full feature set before splitting does not leak test information
    Algorithm 1 lines 20-23 apply ADASYN to C and Y before train/val and train/test splits; no justification is given for why this is safe.
  • ad hoc to paper Equation 4 (F_DB4 = f_high * f_low) is a valid fusion of wavelet components
    The product of high- and low-frequency wavelet components is asserted without derivation or empirical validation.
  • domain assumption The clinical thresholds in Eqs. 15-19 are correct and transferable across datasets
    Thresholds such as PR>200 ms, QRS>120 ms, ST>0.1 mV are taken from literature and assumed to be valid for the datasets used.
  • domain assumption Resampling all signals to 250 Hz and 1500 samples preserves diagnostic information
    Section IV states signals are resampled; this discards information from original 500 Hz or 1000 Hz recordings and is not empirically validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FoundationalECGNet: A Lightweight Foundational Model for ECG-based Multitask Cardiac Analysis." pith.science (2026). https://pith.science/paper/MMYV3SQ4

@misc{pith2026250908961,
  author       = {Pith},
  title        = {Pith review of: FoundationalECGNet: A Lightweight Foundational Model for ECG-based Multitask Cardiac Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MMYV3SQ4}},
  note         = {Machine review of arXiv:2509.08961}
}
read the original abstract

Cardiovascular diseases (CVDs) remain a leading cause of mortality worldwide, underscoring the importance of accurate and scalable diagnostic systems. Electrocardiogram (ECG) analysis is central to detecting cardiac abnormalities, yet challenges such as noise, class imbalance, and dataset heterogeneity limit current methods. To address these issues, we propose FoundationalECGNet, a foundational framework for automated ECG classification. The model integrates a dual-stage denoising by Morlet and Daubechies wavelets transformation, Convolutional Block Attention Module (CBAM), Graph Attention Networks (GAT), and Time Series Transformers (TST) to jointly capture spatial and temporal dependencies in multi-channel ECG signals. FoundationalECGNet first distinguishes between Normal and Abnormal ECG signals, and then classifies the Abnormal signals into one of five cardiac conditions: Arrhythmias, Conduction Disorders, Myocardial Infarction, QT Abnormalities, or Hypertrophy. Across multiple datasets, the model achieves a 99% F1-score for Normal vs. Abnormal classification and shows state-of-the-art performance in multi-class disease detection, including a 99% F1-score for Conduction Disorders and Hypertrophy, as well as a 98.9% F1-score for Arrhythmias. Additionally, the model provides risk level estimations to facilitate clinical decision-making. In conclusion, FoundationalECGNet represents a scalable, interpretable, and generalizable solution for automated ECG analysis, with the potential to improve diagnostic precision and patient outcomes in healthcare settings. We'll share the code after acceptance.

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. An Explainable Vision-Language Model Framework with Adaptive PID-Tversky Loss for Lumbar Spinal Stenosis Diagnosis

    cs.CV 2026-04 unverdicted novelty 6.0 of 10

    A VLM framework with spatial patch cross-attention and adaptive PID-Tversky loss reports 90.69% classification accuracy, 0.9512 Dice score, and 92.80 CIDEr for LSS diagnosis plus automated report generation.

Reference graph

Works this paper leans on

66 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Large-scale Training of Foundation Mod- els for Wearable Biosignals

    Salar Abbaspourazad et al. “Large-scale Training of Foundation Mod- els for Wearable Biosignals”. In: ICLR. 2024

  2. [2]

    A novel multi-class imbalanced EEG signals classifi- cation based on the adaptive synthetic sampling (ADASYN) approach

    Adi Alhudhaif. “A novel multi-class imbalanced EEG signals classifi- cation based on the adaptive synthetic sampling (ADASYN) approach”. In: PeerJ Computer Science (2021)

  3. [3]

    Early diagnosis of cardiovascular diseases in the era of artificial intelligence: An in-depth review

    Naiela E Almansouri et al. “Early diagnosis of cardiovascular diseases in the era of artificial intelligence: An in-depth review”. In: Cureus 16.3 (2024)

  4. [4]

    Deep learning for ECG Arrhythmia detection and classification: an overview of progress for period 2017–2023

    Yaqoob Ansari et al. “Deep learning for ECG Arrhythmia detection and classification: an overview of progress for period 2017–2023”. In: Frontiers in Physiology (2023)

  5. [5]

    Interpretable Deep Learning Models for Arrhythmia Classification Based on ECG Signals Using PTB-X Dataset

    Ahmed E Mansour Atwa et al. “Interpretable Deep Learning Models for Arrhythmia Classification Based on ECG Signals Using PTB-X Dataset”. In: Diagnostics (2025)

  6. [6]

    A comprehensive review of AI-Based detection of Arrhythmia using Electrocardiogram (ECG)

    Ahtisham Ayyub, Christos Politis, and Muhammad Arslan Usman. “A comprehensive review of AI-Based detection of Arrhythmia using Electrocardiogram (ECG)”. In: Computers in Biology and Medicine (2025)

  7. [7]

    An efficient ECG signals denoising technique based on the combination of particle swarm optimisation and wavelet transform

    Abdallah Azzouz et al. “An efficient ECG signals denoising technique based on the combination of particle swarm optimisation and wavelet transform”. In: Heliyon (2024)

  8. [8]

    Implementation of ECG signal processing and analysis techniques in digital signal processor based system

    D Balasubramaniam and D Nedumaran. “Implementation of ECG signal processing and analysis techniques in digital signal processor based system”. In: 2009 IEEE International Workshop on Medical Measurements and Applications . IEEE. 2009

Show all 66 references
  1. [9]

    ExChanGeAI: An End-to-End Platform and Efficient Foundation Model for Electrocardiogram Analysis and Fine- tuning

    Lucas Bickmann et al. “ExChanGeAI: An End-to-End Platform and Efficient Foundation Model for Electrocardiogram Analysis and Fine- tuning”. In: arXiv preprint arXiv:2503.13570 (2025)

  2. [10]

    A literature review: ECG-based models for arrhyth- mia diagnosis using artificial intelligence techniques

    Abir Boulif et al. “A literature review: ECG-based models for arrhyth- mia diagnosis using artificial intelligence techniques”. In: Bioinformat- ics and Biology Insights 17 (2023), p. 11779322221149600

  3. [11]

    Toward automated feature extraction for deep learning classification of electrocardiogram signals

    Fatima Sajid Butt et al. “Toward automated feature extraction for deep learning classification of electrocardiogram signals”. In: IEEE Access (2022)

  4. [12]

    Basic Paediatric ECG Interpretation Principles

    Silvia Cesini et al. “Basic Paediatric ECG Interpretation Principles”. In: New Concepts in ECG Interpretation. Springer, 2018, pp. 195–209

  5. [13]

    AF classification from a short single lead ECG recording: The PhysioNet/computing in cardiology challenge 2017

    Gari D Clifford et al. “AF classification from a short single lead ECG recording: The PhysioNet/computing in cardiology challenge 2017”. In: 2017 computing in cardiology (CinC) . IEEE. 2017

  6. [14]

    Basic electrocardiography: normal and ab- normal ECG patterns

    Antoni Bayés De Luna. Basic electrocardiography: normal and ab- normal ECG patterns . John Wiley & Sons, 2008

  7. [15]

    Fusion of edge detection and graph neural networks to classifying electrocardiogram signals

    Linh T. Duong et al. “Fusion of edge detection and graph neural networks to classifying electrocardiogram signals”. In: Expert Systems with Applications (2023)

  8. [16]

    MedalCare-XL: 16,900 healthy and pathological synthetic 12 lead ECGs from electrophysiological simulations

    Karli Gillette et al. “MedalCare-XL: 16,900 healthy and pathological synthetic 12 lead ECGs from electrophysiological simulations”. In: Scientific Data (2023)

  9. [17]

    Using domain adaptation for classification of healthy and disease conditions from mobile-captured images of standard 12-lead electrocardiograms

    Vadim Gliner, Vladislav Makarov, Artur I. Avetisyan, et al. “Using domain adaptation for classification of healthy and disease conditions from mobile-captured images of standard 12-lead electrocardiograms”. In: Scientific Reports (2023)

  10. [18]

    Components of a new research resource for complex physiologic signals

    Ary Goldberger et al. “Components of a new research resource for complex physiologic signals”. In: PhysioNet (2000)

  11. [19]

    Artifact removal from ECG signals using online recursive independent component analysis

    K Gunasekaran, VD Ambeth Kumar, et al. “Artifact removal from ECG signals using online recursive independent component analysis”. In: Journal of Computational Mathematics and Data Science (2024)

  12. [20]

    A comprehensive review on efficient artificial intelligence models for classification of abnormal cardiac rhythms using electrocardiograms

    Utkarsh Gupta et al. “A comprehensive review on efficient artificial intelligence models for classification of abnormal cardiac rhythms using electrocardiograms”. In: Heliyon (2024)

  13. [21]

    Foundation Models in Electrocardiogram: A Review

    Yu Han et al. Foundation Models in Electrocardiogram: A Review . 2024

  14. [22]

    Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network

    Awni Y Hannun et al. “Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network”. In: Nature medicine (2019)

  15. [23]

    Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network

    Awni Y . Hannun et al. “Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network”. In: Nature Medicine (2019)

  16. [24]

    A transformer-based deep neural network for arrhythmia detection using continuous ECG signals

    Rui Hu, Jie Chen, and Li Zhou. “A transformer-based deep neural network for arrhythmia detection using continuous ECG signals”. In: Computers in Biology and Medicine (2022)

  17. [25]

    Using AUC and accuracy in evaluating learning algorithms

    Jin Huang and Charles X Ling. “Using AUC and accuracy in evaluating learning algorithms”. In: IEEE Transactions on knowledge and Data Engineering (2005)

  18. [26]

    The Applications of Deep Learning in ECG Classification for Disease Diagnosis: A Systematic Review and Meta- Data Analysis

    Muhammad Khalid et al. “The Applications of Deep Learning in ECG Classification for Disease Diagnosis: A Systematic Review and Meta- Data Analysis”. In: Engineering Journal (2024)

  19. [27]

    A novel hybrid CNN-transformer model for arrhythmia detection without R-peak identification using stockwell transform

    Donghyeon Kim et al. “A novel hybrid CNN-transformer model for arrhythmia detection without R-peak identification using stockwell transform”. In: Scientific Reports (2025). 10 FOUNDA TIONALECGNET

  20. [28]

    Comparative study of ECG signal denoising by wavelet thresholding in empirical and variational mode decomposition domains

    Saeid Lahmiri. “Comparative study of ECG signal denoising by wavelet thresholding in empirical and variational mode decomposition domains”. In: Healthcare Technology Letters (2014)

  21. [29]

    DiffuSETS: 12-Lead ECG generation conditioned on clinical text reports and patient-specific information

    Yongfan Lai et al. “DiffuSETS: 12-Lead ECG generation conditioned on clinical text reports and patient-specific information”. In: Patterns (2025)

  22. [30]

    Clinical tests: sensitivity and specificity

    Abdul Ghaaliq Lalkhen and Anthony McCluskey. “Clinical tests: sensitivity and specificity”. In: Continuing education in anaesthesia, critical care & pain (2008)

  23. [31]

    ECG Signal Denoising Method Based on Disentangled Autoencoder

    Haicai Lin, Ruixia Liu, and Zhaoyang Liu. “ECG Signal Denoising Method Based on Disentangled Autoencoder”. In: Electronics (2023)

  24. [32]

    Feature extraction based on Morlet wavelet and its application for mechanical fault diagnosis

    Jing Lin and Liangsheng Qu. “Feature extraction based on Morlet wavelet and its application for mechanical fault diagnosis”. In: Journal of sound and vibration (2000)

  25. [33]

    Multi-dimensional signal processing for non-linear structural dynamics

    H Liu and T Vinh. “Multi-dimensional signal processing for non-linear structural dynamics”. In: Mechanical Systems and Signal Processing (1991)

  26. [34]

    Design and use of a Denoising Convolutional Autoencoder for reconstructing electrocardiogram signals at super resolution

    Ugo Lomoio et al. “Design and use of a Denoising Convolutional Autoencoder for reconstructing electrocardiogram signals at super resolution”. In: Artificial Intelligence in Medicine (2025)

  27. [35]

    Current and future use of artificial intelligence in electrocardiography

    Manuel Martínez-Sellés and Manuel Marina-Breysse. “Current and future use of artificial intelligence in electrocardiography”. In: Journal of Cardiovascular Development and Disease (2023)

  28. [36]

    ECG arrhythmias classification based on deep learning methods and transfer learning technique

    Samira Mavaddati. “ECG arrhythmias classification based on deep learning methods and transfer learning technique”. In: Biomedical Signal Processing and Control (2025)

  29. [37]

    ECG-FM: An Open Electrocardiogram Founda- tion Model

    Kaden McKeen et al. ECG-FM: An Open Electrocardiogram Founda- tion Model. 2025

  30. [38]

    Deep Generative Models: The winning key for large and easily accessible ECG datasets?

    Giuliana Monachino et al. “Deep Generative Models: The winning key for large and easily accessible ECG datasets?” In: Computers in biology and medicine (2023)

  31. [39]

    Data oversampling and imbalanced datasets: an investigation of performance for machine learning and feature engi- neering. J Big Data. 2024; 11: 87

    M Mujahid et al. “Data oversampling and imbalanced datasets: an investigation of performance for machine learning and feature engi- neering. J Big Data. 2024; 11: 87”. In: IEEE Access (2024)

  32. [40]

    Transfer learning in ECG diagnosis: Is it effective?

    Cuong V . Nguyen and Cuong D. Do. “Transfer learning in ECG diagnosis: Is it effective?” In: PLOS ONE (2025)

  33. [41]

    Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation

    David MW Powers. “Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation”. In: arXiv preprint arXiv:2010.16061 (2020)

  34. [42]

    Robustness of Deep Learning models in electro- cardiogram noise detection and classification

    Saifur Rahman et al. “Robustness of Deep Learning models in electro- cardiogram noise detection and classification”. In: Computer Methods and Programs in Biomedicine (2024)

  35. [43]

    Pentti M Rautaharju, Borys Surawicz, and Leonard S Gettes. “AHA/ACCF/HRS recommendations for the standardization and in- terpretation of the electrocardiogram: part IV: the ST segment, T and U waves, and the QT interval a scientific statement from the American Heart Associatio...

  36. [44]

    A Novel Deep Learning based Gated Recurrent Unit with Extreme Learning Machine for Electrocardiogram (ECG) Signal Recognition

    Clement Virgeniya S. and Ramaraj E. “A Novel Deep Learning based Gated Recurrent Unit with Extreme Learning Machine for Electrocardiogram (ECG) Signal Recognition”. In: Biomedical Signal Processing and Control (2021)

  37. [45]

    Risk factors of deaths related to cardiovascular diseases in World Health Organization (WHO) member countries

    Bayram ¸ Sahin and Gülnur ˙Ilgün. “Risk factors of deaths related to cardiovascular diseases in World Health Organization (WHO) member countries”. In: Health & Social Care in the Community (2022)

  38. [46]

    A Systematic Review of ECG Arrhythmia Clas- sification: Adherence to Standards, Fair Evaluation, and Embedded Feasibility

    Guilherme Silva et al. A Systematic Review of ECG Arrhythmia Clas- sification: Adherence to Standards, Fair Evaluation, and Embedded Feasibility. 2025

  39. [47]

    Unveiling person- ality traits through Bangla speech using Morlet wavelet transformation and BiG

    Md Sajeebul Islam Sk and Md Golam Rabiul Alam. “Unveiling person- ality traits through Bangla speech using Morlet wavelet transformation and BiG”. In: Natural Language Processing Journal (2024)

  40. [48]

    CREMA: A Contrastive Regularized Masked Autoen- coder for Robust ECG Diagnostics across Clinical Domains

    Junho Song et al. CREMA: A Contrastive Regularized Masked Autoen- coder for Robust ECG Diagnostics across Clinical Domains . 2025. arXiv: 2407.07110

  41. [49]

    Investigating advantages and disadvantages of the analysis of a geometrical surface structure with the use of Fourier and wavelet transform

    Krzysztof St˛ epie ´n, Włodzimierz Makieła, and Stanisław Adamczak. “Investigating advantages and disadvantages of the analysis of a geometrical surface structure with the use of Fourier and wavelet transform”. In: Metrology and Measurement Systems (2010)

  42. [50]

    Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL

    Nils Strodthoff et al. “Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL”. In: IEEE Journal of Biomedical and Health Informatics (2021)

  43. [51]

    Deep learning for ECG analysis: Benchmarks and insights from PTB-XL

    Nils Strodthoff et al. “Deep learning for ECG analysis: Benchmarks and insights from PTB-XL”. In: IEEE journal of biomedical and health informatics (2020)

  44. [52]

    Class-driven graph attention network for multi-label time series classification in mobile health digital twins

    Le Sun et al. “Class-driven graph attention network for multi-label time series classification in mobile health digital twins”. In: IEEE Journal on Selected Areas in Communications (2023)

  45. [53]

    Foundation model of ECG diagnosis: Diag- nostics and explanations of any form and rhythm on ECG

    Yuanyuan Tian et al. “Foundation model of ECG diagnosis: Diag- nostics and explanations of any form and rhythm on ECG”. In: Cell Reports Medicine 5.12 (2024)

  46. [54]

    Drug-induced arrhythmias: a scientific state- ment from the American Heart Association

    James E Tisdale et al. “Drug-induced arrhythmias: a scientific state- ment from the American Heart Association”. In: Circulation (2020)

  47. [55]

    Generalized Daubechies wavelet families

    Cédric V onesch, Thierry Blu, and Michael Unser. “Generalized Daubechies wavelet families”. In: IEEE transactions on signal pro- cessing (2007)

  48. [56]

    Interactive ECG annotation: An artificial intelli- gence method for smart ECG manipulation

    Haiyan Wang et al. “Interactive ECG annotation: An artificial intelli- gence method for smart ECG manipulation”. In: Information Sciences (2021)

  49. [57]

    AnyECG: Foundational Models for Multitask Cardiac Analysis in Real-World Settings

    Yue Wang et al. “AnyECG: Foundational Models for Multitask Cardiac Analysis in Real-World Settings”. In: arXiv preprint arXiv:2411.17711 (2024)

  50. [58]

    Transfer learning for ECG classification

    Kuba Weimann and Tim O. F. Conrad. “Transfer learning for ECG classification”. In: Scientific Reports (2021)

  51. [59]

    Deep learning and electrocardiography: systematic review of current techniques in cardiovascular disease diag- nosis and management

    Zhenyan Wu and Caixia Guo. “Deep learning and electrocardiography: systematic review of current techniques in cardiovascular disease diag- nosis and management”. In: BioMedical Engineering OnLine (2025)

  52. [60]

    Denoising ECG signal using Daubechies and Symlet wavelet transform techniques

    Tanuj Yadav and Rajeev Mehra. “Denoising ECG signal using Daubechies and Symlet wavelet transform techniques”. In: Interna- tional Journal of Advanced Research in Computer and Communication Engineering (2016)

  53. [61]

    Clyde W Yancy et al. “2017 ACC expert consensus decision pathway for optimization of heart failure treatment: answers to 10 pivotal issues about heart failure with reduced ejection fraction: a report of the American College of Cardiology Task Force on Expert Consensus Decision...

  54. [62]

    Brant: Foundation model for intracranial neu- ral signal

    Daoze Zhang et al. “Brant: Foundation model for intracranial neu- ral signal”. In: Advances in Neural Information Processing Systems (2023)

  55. [63]

    Opportunities and challenges of noise interference suppression algorithms for dynamic ECG signals in wearable devices: A review

    Juya Zhang et al. “Opportunities and challenges of noise interference suppression algorithms for dynamic ECG signals in wearable devices: A review”. In: Measurement (2025), p. 117067

  56. [64]

    Lumbar spine localisation method based on feature fusion

    Yonghong Zhang et al. “Lumbar spine localisation method based on feature fusion”. In: CAAI Transactions on Intelligence Technology (2023)

  57. [65]

    Transforming ECG Diagnosis:An In-depth Review of Transformer-based DeepLearning Models in Cardiovascular Disease Detection

    Zibin Zhao. Transforming ECG Diagnosis:An In-depth Review of Transformer-based DeepLearning Models in Cardiovascular Disease Detection. 2023. eprint: 2306.01249

  58. [66]

    A cascaded multi-stage framework for automatic detection and segmentation of pulmonary nodules in developing coun- tries

    Zhixun Zhou et al. “A cascaded multi-stage framework for automatic detection and segmentation of pulmonary nodules in developing coun- tries”. In: IEEE Journal of Biomedical and Health Informatics (2022)

Pith tools

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