Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

A Novel Data Augmentation Strategy for Robust Deep Learning Classification of Biomedical Time-Series Data: Application to ECG and EEG Analysis

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that concatenating seven augmented copies of each ECG or EEG signal into a single input lets one ResNet-plus-attention model reach 99.96–100% accuracy on three benchmarks.

desk verdict Testable augmentation idea, but missing split protocol and ablation make the SOTA claims unverified; deserves a critical referee, not rejection. read the letter →

arxiv 2507.12645 v1 pith:PUSNAJPL submitted 2025-07-16 eess.SP cs.AIcs.LG

Mohammed Guhdar , Ramadhan J. Mstafa , Abdulhakeem O. Mohammed This is my paper · ORCID
classification eess.SPcs.AIcs.LG
keywords biomedicaltime-seriesclassificationECGEEGdataaugmentationtime-domainconcatenationResNetattentionmechanismfocalloss
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 claims that a single deep-learning architecture can classify both EEG and ECG signals at the top of published accuracy tables by changing how augmented training data are presented. Instead of training on each augmented copy separately, it concatenates seven versions of every signal—original, noisy, scaled, circularly shifted, time-warped, cutout, and amplitude-jittered—into one long input, forcing the model to learn features that survive all those distortions at once. With a ResNet-style convolutional backbone, a channel-wise attention gate, and focal loss, the model reports 99.96% accuracy on UCI Seizure EEG, 99.78% on MIT-BIH arrhythmia, and 100% on PTB diagnostic ECG. A sympathetic reader would care because the method is deliberately simple, unified across two signal modalities, and light enough (about 130 MB and roughly 10 ms per sample) to run on modest hardware, if the results hold under honest evaluation.

What carries the argument

The load-bearing mechanism is the concatenation operation $s_{\text{concat}} = [s_{\text{orig}}, s_{\text{noisy}}, s_{\text{scaled}}, s_{\text{shifted}}, s_{\text{warped}}, s_{\text{cutout}}, s_{\text{jitter}}]$, which merges seven distorted copies of each preprocessed signal into one input. This is paired with a three-block ResNet backbone (kernel size 5 residual blocks, initial convolution of kernel 15), an attention block that computes channel-wise gates via a $C \to C/8 \to C$ bottleneck with sigmoid activation, and a two-layer classifier with dropout 0.6. Focal loss with default focusing parameter $\gamma=2$ and AdamW weight decay handle imbalance and regularization. The concatenation does the main work: it forces the network to learn transformation-invariant features from a single forward pass, which the paper treats as a stronger form of implicit data augmentation and regularization.

What would settle it

Re-run the same preprocessing, augmentation, and architecture on the three datasets with a subject-exclusive split protocol—assigning all segments or beats of each patient either wholly to training or wholly to test—and compare accuracy and F1; if the 100% PTB and 99.96% UCI results drop materially under that protocol, the claimed generalization to new patients is not supported.

Watch

Extended reading notes

Core claim

The central claim is that time-domain concatenation of multiple augmented variants—the original signal plus noisy, scaled, shifted, time-warped, cutout, and amplitude-jittered copies, joined end-to-end into a single training instance—makes a ResNet-plus-attention classifier simultaneously robust and accurate across heterogeneous biomedical time-series. The authors argue this creates a stronger regularizer than training on augmented examples separately, because the network must extract features present in the original signal and all its transformed versions in one pass. Evaluated after wavelet denoising, baseline removal, and standardization, the model achieves 99.96% accuracy and 0.9989 F1 on the UCI seizure EEG set, 99.78% and 0.9937 on MIT-BIH, and 100% with F1 of 1.0 on PTB, and the authors position these as state-of-the-art results on those benchmarks. They further claim the design handles class imbalance without oversampling, thanks to focal loss combined with the augmentation.

Load-bearing premise

The evaluation never states how segments are assigned to training, validation, and test sets; if the assignment is random at the segment level rather than subject-disjoint, the same patient's data can appear on both sides of the split, which alone could produce the reported near-perfect accuracies.

Editorial extensions

If this is right

  • If the results reproduce, a single architecture can serve as a common ECG/EEG classifier, reducing the need for modality-specific preprocessing and feature engineering.
  • The concatenation strategy could replace or supplement oversampling in imbalanced clinical datasets, since the paper reports high minority-class performance on MIT-BIH without explicit oversampling.
  • The reported 100% on PTB and 99.96% on UCI would make those benchmarks close to saturated under this protocol, shifting the field to harder splits or multi-channel inputs.
  • At roughly 130 MB and 10 ms per sample, the model is small enough for wearable or low-end deployment, enabling on-device screening of cardiac and neurological signals.

Reading between the lines

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

  • The paper never states whether train/validation/test splits are subject-disjoint; if segments are split randomly, the same patient's recordings can appear in both training and test, and the near-perfect scores would partly reflect patient identity rather than generalization to new people. This is my inference, not a claim the paper makes.
  • The augmentation multiplies input length sevenfold, so the reported 10 ms inference and roughly 130 MB memory figure refer to the full ensemble; applying the same concatenation to multi-channel 12-lead ECG would increase compute and memory proportionally, a limitation the authors themselves acknowledge.
  • The concatenation idea is generic: it should transfer to other time-series domains such as speech, motion, or vibration, as long as the chosen distortions match the physical noise sources, and a direct test would be whether accuracy gains persist under explicit subject-disjoint evaluation.
  • The comparison table mostly lists earlier published accuracies; a direct reproduction on identical split protocols would settle whether the improvement over the best prior method (99.83% on PTB) reflects the method or differences in evaluation setup.
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

5 major / 5 minor

Summary. The paper proposes a unified deep learning framework for ECG and EEG classification, combining a ResNet-based CNN with an attention mechanism and a data augmentation strategy that concatenates seven augmented variants of each signal in the time domain. The authors report state-of-the-art accuracies of 99.96% on UCI Seizure EEG, 99.78% on MIT-BIH Arrhythmia, and 100% on PTB Diagnostic ECG. They also estimate ~130 MB memory footprint and ~10 ms inference time per sample, suggesting wearable deployment.

Significance. If the reported results are valid, the framework would be a practically attractive unified solution for binary ECG/EEG classification, with a simple augmentation scheme and modest resource requirements. The paper contains useful details on preprocessing, focal loss, and regularization. However, the evaluation has serious gaps: no train/test split protocol is described, the inference-time input construction is unspecified, the ensemble is not defined, and no ablation isolates the augmentation contribution. The near-perfect accuracy values therefore cannot be assessed for subject leakage or overfitting. The significance is conditional on the authors providing a rigorous, subject-disjoint evaluation and supporting ablations.

major comments (5)
  1. [Section 3.1 and Section 6.1] The manuscript never states how segments are assigned to training, validation, and test sets. Since the PTB database has 290 subjects (Table 3) and the UCI dataset comprises recordings from individuals, a random segment-level split would place same-subject segments in both training and test, which can artificially inflate accuracy to the near-perfect levels reported. Please specify the exact split protocol and provide a subject-disjoint evaluation, with test set sizes and class distributions.
  2. [Section 4 and Section 5] The training input is the concatenation s_concat (Eq. 12), but the paper does not describe how test inputs are constructed. If test segments are presented as single preprocessed signals, the input length differs from the trained model; if they are also concatenated with randomly augmented variants, the test procedure is stochastic and the reported metrics are not well defined. The manuscript must specify the exact inference-time input and how the reported test metrics were computed.
  3. [Table 4 and Section 6.1] The caption and the resource-utilization paragraph refer to an 'ensemble implementation comprising five model instances,' but Section 5 describes only a single model forward pass (Eq. 23). The ensemble construction, training procedure, and combination rule are never specified. Please clarify whether Table 4 reports the ensemble or a single model, and provide the full ensemble details.
  4. [Section 4 and Section 6.5] The paper's central contribution is the time-domain concatenation augmentation strategy, yet no ablation isolates its effect. The experiments should compare (i) no augmentation, (ii) the same augmented variants used as independent training samples rather than concatenated, and (iii) the proposed concatenation, on at least one dataset. Without this, the attribution of the state-of-the-art results to the augmentation strategy is unsupported.
  5. [Section 6.2-6.4] All reported metrics are from a single evaluation, with no error bars, confidence intervals, or number of runs. Given the ceiling-level values (accuracy 1.0 and F1 1.0 on PTB, 0.9996 on UCI), the results should be reported over multiple seeds with bootstrap confidence intervals, along with the exact test set sizes.
minor comments (5)
  1. [Equation (7)] The notation N(0, 0.012) conflicts with the parenthetical 'standard deviation 0.01'; use N(0, (0.01)^2) or otherwise clarify the variance.
  2. [Table 1] Labeling the 178 features per segment as 'Number of Channels' is misleading; the UCI dataset provides 178 values per one-second segment, not 178 channels.
  3. [Section 5.2] The variable names in Eqs. (13)-(18) are inconsistent (e.g., h1 is reused); please renumber or rename the intermediate activations for clarity.
  4. [Section 7] The conclusion mentions 'dynamic learning rate adjustment' but Section 5.6 does not describe a learning-rate scheduler; either add the description or remove the phrase.
  5. [Section 6.1] The resource analysis uses 'GPU FLOPs per Second' in Eq. (33) but the text discusses CPU; correct this inconsistency.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the reported accuracies are empirical outcomes from training on fixed public benchmarks, and the only self-citation [40] is a non-load-bearing baseline row.

full rationale

The paper's central claim is an empirical accuracy claim, not a quantity derived from an input. The augmentation transformation is a fixed concatenation (Eq. 12) of hand-specified variants (Eqs. 6-11), with parameters such as noise variance 0.012, warp factor 0.2, cutout length 50, and jitter 0.05 stated in advance; none of these are fitted to test labels. Metrics in Eqs. 26-31 are standard definitions applied to predictions, so no reported number reduces to a fitted parameter or to the labels by construction. The architecture is a standard ResNet+attention+focal-loss pipeline; the comparative SOTA claim is supported by testing on public benchmark datasets (UCI, MIT-BIH, PTB), which is external evidence. The only self-reference is row [40] in Table 5, the authors' earlier 1D-CNN+attention model, used as one of many baselines; it does not supply any assumption, theorem, or fitted value to the proposed model, so it is not load-bearing. Separately, the paper omits a patient-disjoint train/test split description (Section 3.1/6.1), which is a validity risk because same-subject segments could leak between folds; that concern is about experimental protocol, not circularity. The conclusion's stated scalability limitation for 12-lead input is likewise a limitation, not a circular step. Accordingly, no circular step is present; the minor self-citation and missing split protocol are noted without raising the score above the low end.

Assumptions & free parameters 9 free parameters · 3 assumptions · 0 invented entities

No constants were fitted to the reported test labels. All listed quantities are hand-selected hyperparameters or preprocessing choices that the central claim depends on; none has an independent justification beyond prior practice. The paper introduces no new physical or mathematical entities.

free parameters (9)
  • Noise variance for noisy augmentation = 0.012 (text also says standard deviation 0.01)
    Hand-chosen in Eq. (7) to simulate sensor noise; the paper gives inconsistent notation.
  • Scale factor for scaled augmentation = 1.2
    Eq. (8), hand-chosen amplitude multiplier.
  • Circular shift amount = 10 samples
    Eq. (9), hand-chosen time shift.
  • Time warp factor = 0.2
    Eq. (10), controls maximum warping.
  • Cutout segment count and length = 2 segments of length 50
    Section 4 defaults, hand-chosen.
  • Jitter standard deviation = 0.05
    Eq. (11), hand-chosen noise level.
  • Dropout probability = 0.6
    Section 5.4, hand-chosen.
  • Weight decay coefficient = 0.01
    Section 5.6.2, hand-chosen.
  • Focal loss gamma = 2 (alpha implicitly 1)
    Eq. (25), standard values but still hand-chosen.
assumptions (3)
  • domain assumption The benchmark datasets are evaluated under a split that prevents subject-level data leakage.
    Section 6.1 reports validation results without ever specifying how segments were assigned to train/validation/test sets, so the validity of the accuracies depends on an assumed clean split.
  • domain assumption Time-domain concatenation of augmented variants is a stronger regularizer than training on augmented examples separately.
    Section 4 asserts this claim ('a much stronger form of regularization') without an ablation or theoretical support.
  • domain assumption Wavelet denoising with db4 at level 4 and median-filter baseline removal preserve the discriminative signal content.
    Section 3.2 applies these preprocessing steps as standard practice, but their effect on classification is not isolated or validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Novel Data Augmentation Strategy for Robust Deep Learning Classification of Biomedical Time-Series Data: Application to ECG and EEG Analysis." pith.science (2026). https://pith.science/paper/PUSNAJPL

@misc{pith2026250712645,
  author       = {Pith},
  title        = {Pith review of: A Novel Data Augmentation Strategy for Robust Deep Learning Classification of Biomedical Time-Series Data: Application to ECG and EEG Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PUSNAJPL}},
  note         = {Machine review of arXiv:2507.12645}
}
read the original abstract

The increasing need for accurate and unified analysis of diverse biological signals, such as ECG and EEG, is paramount for comprehensive patient assessment, especially in synchronous monitoring. Despite advances in multi-sensor fusion, a critical gap remains in developing unified architectures that effectively process and extract features from fundamentally different physiological signals. Another challenge is the inherent class imbalance in many biomedical datasets, often causing biased performance in traditional methods. This study addresses these issues by proposing a novel and unified deep learning framework that achieves state-of-the-art performance across different signal types. Our method integrates a ResNet-based CNN with an attention mechanism, enhanced by a novel data augmentation strategy: time-domain concatenation of multiple augmented variants of each signal to generate richer representations. Unlike prior work, we scientifically increase signal complexity to achieve future-reaching capabilities, which resulted in the best predictions compared to the state of the art. Preprocessing steps included wavelet denoising, baseline removal, and standardization. Class imbalance was effectively managed through the combined use of this advanced data augmentation and the Focal Loss function. Regularization techniques were applied during training to ensure generalization. We rigorously evaluated the proposed architecture on three benchmark datasets: UCI Seizure EEG, MIT-BIH Arrhythmia, and PTB Diagnostic ECG. It achieved accuracies of 99.96%, 99.78%, and 100%, respectively, demonstrating robustness across diverse signal types and clinical contexts. Finally, the architecture requires ~130 MB of memory and processes each sample in ~10 ms, suggesting suitability for deployment on low-end or wearable devices.

Figures

Figures reproduced from arXiv: 2507.12645 by the authors.

Figure 1
Figure 1. ECG Signal before and after applying data augmentation [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. EEG Signal before and after applying data augmentation [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Data Augmentation dimension to 1, resulting in a feature vector of size 512. This vector is then processed by an atten￾tion block (attention) to weigh feature importance before the final classification by the classifier. The classifier consists of two fully connected layers with ReLU activation and dropout in between, producing the output probabilities. 5.2. ResNet Backbone The ResNet backbone comprises three sequen… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Flowchart of the Proposed Model CSI = TP TP + FN + FP (30) MCC = TP × TN − FP × FN √ (TP + FP)(TP + FN)(TN + FP)(TN + FN) (31) 6. Results and Discussion 6.1. Performance Validation and Robustness This section presents a comprehensive analysis of the experimental result…
Figure 5
Figure 5. Figure 5: Training and Validation Accuracy on the PTB Diagnostic ECG Dataset [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: Confusion Matrix on the PTB Diagnostic ECG Dataset [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: Training and Validation Accuracy on the MIT-BIH Arrhythmia Dataset [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Confusion Matrix on the MIT-BIH Arrhythmia Dataset [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Training and Validation Accuracy on the UCI Seizure EEG Dataset [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Confusion Matrix on the UCI Seizure EEG Dataset [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. A Comprehensive Inference-Time Augmentation Framework in Physiological Signals: Application to PPG-Based AF Detection

    cs.LG 2026-06 unverdicted novelty 6.0 of 10

    A unified inference-time augmentation framework with 13 methods and Bayesian-optimized parameters improves AUROC up to 8.5% and reduces false positives in PPG-based AF detection across five datasets.

Reference graph

Works this paper leans on

51 extracted references · 50 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gogan, O

    A. Gogan, O. Potre, V .-F. Avram, M. Andor, F. Caruntu, B. Timar, Cardiac autonomic neu- ropathy in diabetes mellitus: Pathogenesis, epidemiology, diagnosis and clinical implications: A narrative review, Journal of Clinical Medicine 14 (3) (2025) 671

    work page 2025

  2. [2]

    Brito, M

    C. Brito, M. B. Falc ˜ao, M. d. F. P. M. de Albuquerque, T. Cerqueira-Silva, M. G. Teixeira, R. F. d. O. Franca, Chikungunya: From hypothesis to evidence of increased severe disease and fatalities, Viruses 17 (1) (2025) 62

    work page 2025

  3. [3]

    A. Li, H. Li, G. Yuan, Continual learning with deep neural networks in physiological signal data: A survey, in: Healthcare, V ol. 12, MDPI, 2024, p. 155

    work page 2024

  4. [4]

    Kolhar, A

    M. Kolhar, A. M. Al Rajeh, Deep learning hybrid model ecg classification using alexnet and parallel dual branch fusion network model, Scientific Reports 14 (1) (2024) 26919

    work page 2024

  5. [5]

    Miljkovi ´c, N

    N. Miljkovi ´c, N. B. Popovi ´c, J. Sodnik, Electrogastrogram signal processing: Techniques and challenges with application for simulator sickness assessment, in: Biomedical Signal Processing, CRC Press, 2024, pp. 62–89

    work page 2024

  6. [6]

    Bennia, S

    F. Bennia, S. Moussaoui, M. C. Boutalbi, N. Messaoudi, Comparative study between emd, eemd, and ceemdan based on de-noising bioelectric signals, in: 2024 8th International Con- ference on Image and Signal Processing and their Applications (ISPA), IEEE, 2024, pp. 1–6

    work page 2024

  7. [7]

    Zhang, J

    B. Zhang, J. Wang, Z. Ye, L. Zhou, D. Xiong, X. Wang, L. Guo, Jassnet: Heart and lung sound separation network based on joint attention mechanism and semi-supervised learning, Biomedical Signal Processing and Control 104 (2025) 107525

    work page 2025

  8. [8]

    B. Li, W. Wei, A. Ferreira, S. Tan, Rest-net: Diverse activation modules and parallel subnets- based cnn for spatial image steganalysis, IEEE Signal Processing Letters 25 (5) (2018) 650– 654

    work page 2018

Show all 51 references
  1. [9]

    Z. Niu, G. Zhong, H. Yu, A review on the attention mechanism of deep learning, Neurocom- puting 452 (2021) 48–62

    2021

  2. [10]

    Y . Jin, C. Qin, J. Liu, Y . Liu, Z. Li, C. Liu, A novel deep wavelet convolutional neural network for actual ecg signal denoising, Biomedical Signal Processing and Control 87 (2024) 105480

    2024

  3. [11]

    Dayananda, S

    B. Dayananda, S. Owen, A. Kolobaric, J. Chapman, D. Cozzolino, Pre-processing applied to instrumental data in analytical chemistry: a brief review of the methods and examples, Critical Reviews in Analytical Chemistry 54 (8) (2024) 2745–2753

    2024

  4. [12]

    R. G. Andrzejak, K. Lehnertz, F. Mormann, C. Rieke, P. David, C. E. Elger, Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state, Physical Review E 64 (6) (2001) 061907

    2001

  5. [13]

    G. B. Moody, R. G. Mark, The mit-bih arrhythmia database on cd-rom and software for use with it, in: [1990] Proceedings Computers in Cardiology, IEEE, 1990, pp. 185–188. 24

    1990

  6. [14]

    Bousseljot, D

    R.-D. Bousseljot, D. Kreiseler, A. Schnabel, The ptb diagnostic ecg database, physionet. org (2004)

    2004

  7. [15]

    Wasimuddin, K

    M. Wasimuddin, K. Elleithy, A.-S. Abuzneid, M. Faezipour, O. Abuzaghleh, Stages-based ecg signal analysis from traditional signal processing to machine learning approaches: A survey, IEEE Access 8 (2020) 177782–177803

    2020

  8. [16]

    Petmezas, L

    G. Petmezas, L. Stefanopoulos, V . Kilintzis, A. Tzavelis, J. A. Rogers, A. K. Katsaggelos, N. Maglaveras, State-of-the-art deep learning methods on electrocardiogram data: systematic review, JMIR medical informatics 10 (8) (2022) e38454

    2022

  9. [17]

    D. P. Dash, M. Kolekar, C. Chakraborty, M. R. Khosravi, Review of machine and deep learn- ing techniques in epileptic seizure detection using physiological signals and sentiment analy- sis, ACM Transactions on Asian and Low-Resource Language Information Processing 23 (1) (2024) 1–29

    2024

  10. [18]

    P. N. Singh, R. P. Mahapatra, A novel deep learning approach for arrhythmia prediction on ecg classification using recurrent cnn with gwo, International Journal of Information Tech- nology 16 (1) (2024) 577–585

    2024

  11. [19]

    K. Liu, R. Yang, L. Qi, W. Luan, Z. Zhang, Art-net: An attention-based hybrid resnet- transformer network for 12-lead ecg signal classification, in: 2024 IEEE International Con- ference on Systems, Man, and Cybernetics (SMC), IEEE, 2024, pp. 3664–3670

    2024

  12. [20]

    J. Li, W. Shi, Y . Li, An e ffective classification approach for eeg-based motor imagery tasks combined with attention mechanisms, Cognitive Neurodynamics 18 (5) (2024) 2689–2707

    2024

  13. [21]

    J. Lim, Y . Lee, W. Jang, S. Joo, Specialized ecg data augmentation method: leveraging pre- cordial lead positional variability, Biomedical Engineering Letters (2025) 1–12

    2025

  14. [22]

    L. Fan, T. Chen, L. He, Z. Wang, R. Zhang, Gadm: Data augmentation using generative adversarial diffusion model for pulse-based disease identification, Biomedical Signal Pro- cessing and Control 100 (2025) 107005

    2025

  15. [23]

    Altalhan, A

    M. Altalhan, A. Algarni, M. T.-H. Alouane, Imbalanced data problem in machine learning: A review, IEEE Access (2025)

    2025

  16. [24]

    M. M. Al Rahhal, Y . Bazi, H. Almubarak, N. Alajlan, M. Al Zuair, Dense convolutional networks with focal loss and image generation for electrocardiogram classification, IEEE Access 7 (2019) 182225–182237

    2019

  17. [25]

    Y . Liu, Q. Liu, Smote oversampling algorithm based on generative adversarial network, Clus- ter Computing 28 (4) (2025) 271

    2025

  18. [26]

    Powers, Evaluation: From precision, recall and f-factor to roc, informedness, markedness & correlation (tech

    D. Powers, Evaluation: From precision, recall and f-factor to roc, informedness, markedness & correlation (tech. rep.), Adelaide, Australia (2007)

    2007

  19. [27]

    D. M. Powers, Evaluation: from precision, recall and f-measure to roc, informedness, markedness and correlation, arXiv preprint arXiv:2010.16061 (2020). 25

    2020 arXiv

  20. [28]

    G. K. Mbizvo, K. H. Bennett, C. R. Simpson, S. E. Duncan, R. F. Chin, A. J. Larner, Us- ing critical success index or gilbert skill score as composite measures of positive predictive value and sensitivity in diagnostic accuracy studies: Weather forecasting informing epilepsy r...

    2023

  21. [29]

    Chicco, G

    D. Chicco, G. Jurman, The matthews correlation coe fficient (mcc) should replace the roc auc as the standard metric for assessing binary classification, BioData Mining 16 (1) (2023) 4

    2023

  22. [30]

    W. Fan, Y . Si, W. Yang, M. Sun, Imbalanced ecg data classification using a novel model based on active training subset selection and modified broad learning system, Measurement 198 (2022) 111412

    2022

  23. [31]

    Liotto, A

    C. Liotto, A. Petrillo, S. Santini, G. Toscano, V . Tufano, A multiclass cnn cascade model for the clinical detection support of cardiac arrhythmia based on subject-exclusive ecg dataset, Biomedical Engineering Letters 12 (4) (2022) 433–444

    2022

  24. [32]

    Z. Wang, S. Stavrakis, B. Yao, Hierarchical deep learning with generative adversarial network for automatic cardiac diagnosis from ecg signals, Computers in Biology and Medicine 155 (2023) 106641

    2023

  25. [33]

    B.-T. Pham, P. T. Le, T.-C. Tai, Y .-C. Hsu, Y .-H. Li, J.-C. Wang, Electrocardiogram heartbeat classification for arrhythmias and myocardial infarction, Sensors 23 (6) (2023) 2993

    2023

  26. [34]

    Akbar, S

    M. Akbar, S. Nurmaini, R. U. Partan, The deep convolutional networks for the classification of multi-class arrhythmia, Bulletin of Electrical Engineering and Informatics 13 (2) (2024) 1325–1333

    2024

  27. [35]

    A. K. Dwivedi, G. Srivastava, S. Tripathi, N. Pradhan, efusenet: A deep ensemble fusion network for e fficient detection of arrhythmia and myocardial infarction using ecg signals, Multimedia Tools and Applications (2024) 1–32

    2024

  28. [36]

    Zehir, T

    H. Zehir, T. Hafs, S. Daas, Involutional neural networks for ecg spectrogram classification and person identification, International Journal of Signal and Imaging Systems Engineering 13 (1) (2024) 41–53

    2024

  29. [37]

    H. M. Rai, J. Yoo, S. Dashkevych, Gan-skipnet: A solution for data imbalance in cardiac ar- rhythmia detection using electrocardiogram signals from a benchmark dataset, Mathematics 12 (17) (2024) 2693

    2024

  30. [38]

    Vinutha, U

    K. Vinutha, U. Thirunavukkarasu, Prediction of arrhythmia from mit-bih database using sup- port vector machine (svm) and naive bayes (nb) classifiers, in: AIP Conference Proceedings, V ol. 2853, AIP Publishing, 2024, p. 020198

    2024

  31. [39]

    W. Fan, W. Yang, T. Chen, Y . Guo, Y . Wang, Aocbls: A novel active and online learning sys- tem for ecg arrhythmia classification with less labeled samples, Knowledge-Based Systems 304 (2024) 112553. 26

    2024

  32. [40]

    Guhdar, A

    M. Guhdar, A. O. Mohammed, R. J. Mstafa, Advanced deep learning framework for ecg arrhythmia classification using 1d-cnn with attention mechanism, Knowledge-Based Systems (2025) 113301

    2025

  33. [41]

    K. M. Almustafa, Classification of epileptic seizure dataset using di fferent machine learning algorithms, Informatics in Medicine Unlocked 21 (2020) 100444

    2020

  34. [42]

    Nahzat, M

    S. Nahzat, M. Ya ˘gano˘glu, Classification of epileptic seizure dataset using di fferent machine learning algorithms and pca feature reduction technique, Journal of Investigations on Engi- neering and Technology 4 (2) (2021) 47–60

    2021

  35. [43]

    Abenna, M

    S. Abenna, M. Nahid, H. Bouyghf, B. Ouacha, Eeg-based bci: A novel improvement for eeg signals classification based on real-time preprocessing, Computers In Biology And Medicine 148 (2022) 105931

    2022

  36. [44]

    Gupta, U

    D. Gupta, U. Gupta, H. J. Sarma, Functional iterative approach for universum-based primal twin bounded support vector machine to eeg classification (fuptbsvm), Multimedia Tools and Applications 83 (8) (2024) 22119–22151

    2024

  37. [45]

    Kunekar, C

    P. Kunekar, C. Kumawat, V . Lande, S. Lokhande, R. Mandhana, M. Kshirsagar, Compari- son of di fferent machine learning algorithms to classify epilepsy seizure from eeg signals, Engineering Proceedings 59 (1) (2024) 166

    2024

  38. [46]

    A. Omar, T. Abd El-Hafeez, Optimizing epileptic seizure recognition performance with fea- ture scaling and dropout layers, Neural Computing and Applications 36 (6) (2024) 2835– 2852

    2024

  39. [47]

    Kunekar, M

    P. Kunekar, M. K. Gupta, P. Gaur, Detection of epileptic seizure in eeg signals using machine learning and deep learning techniques, Journal of Engineering and Applied Science 71 (1) (2024) 21

    2024

  40. [48]

    Bhadra, P

    R. Bhadra, P. K. Singh, M. Mahmud, Hyepiseid: a hybrid convolutional neural network and gated recurrent unit model for epileptic seizure detection from electroencephalogram signals, Brain Informatics 11 (1) (2024) 21

    2024

  41. [49]

    M. Jain, H. Bhardwaj, A. Srivastav, Bayesian-enhanced eeg signal analysis for epileptic seizure recognition: A 1d-cnn lstm approach, in: 2024 11th International Conference on Re- liability, Infocom Technologies and Optimization (Trends and Future Directions)(ICRITO), IEEE, 202...

    2024

  42. [50]

    Shawly, A

    T. Shawly, A. A. Alsheikhy, Y . Said, A. I. AbuEid, A. A. Alzahrani, A. A. Alshdadi, H. E. Ahmed, Mafbn: Multi-attention forward and backward network to predict epileptic seizure, Biomedical Signal Processing and Control 104 (2025) 107574

    2025

  43. [51]

    Z. Y . Lim, Y . H. Pang, S. Y . Ooi, W. H. Khoh, F. S. Hiew, Mltcn-eeg: metric learning- based temporal convolutional network for seizure eeg classification, Neural Computing and Applications 37 (4) (2025) 2849–2875. 27

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.