Pith. sign in

REVIEW 5 major objections 4 minor 46 references

A Complex UNet Approach for Non-Invasive Fetal ECG Extraction Using Single-Channel Dry Textile Electrodes

T0 review · 5 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A complex-valued U-Net extracts fetal ECG from single-channel dry textile electrode recordings, preserving P, QRS, and T morphology.

desk verdict A useful simulation pipeline and an honest limitations section, but the state-of-the-art claim doesn't survive contact with the PhysioNet benchmark or real dry-electrode data. read the letter →

arxiv 2506.22457 v1 pith:PNYZCLPL submitted 2025-06-16 eess.SP cs.AI

Iulia Orvas , Andrei Radu , Alessandra Galli , Ana Neacsu , Elisabetta Peri This is my paper · ORCID
classification eess.SPcs.AI
keywords fetalECGextractiondrytextileelectrodessingle-channelabdominalcomplex-valuedneuralnetworkU-NetarchitecturespectrogramphaseR-peakdetectionin-silicodataset
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

This paper tries to establish that fetal ECG can be extracted from a single abdominal channel recorded with dry textile electrodes, using a complex-valued U-Net that works on both the magnitude and phase of a spectrogram. The authors argue this is the first method to do so, and that it outperforms real-valued U-Nets, attention U-Nets, and established signal-processing baselines on both simulated and real recordings. If correct, the method would let pregnant women wear a small textile sensor at home and obtain clinically usable fetal heart traces without gel electrodes or hospital visits. The quantitative support comes from a simulated test set where the method reaches PRD 26.2 ± 11.6 and PCC 95.9 ± 5.7 for waveform extraction, and F-score 77.8 ± 18.6 on a standard wet-electrode benchmark dataset.

What carries the argument

The Complex U-Net (CUNet): a U-Net-style network whose encoder and decoder operate on complex-valued spectrograms, splitting each complex convolution into real and imaginary parts, with a trainable diagonal phase-shift layer placed at the first and last layers and a learnable convex combination of Complex ReLU, Georgiou-Koutsougeras, and GroupSort activations. By processing magnitude and phase jointly, the network learns to separate fetal ECG from maternal ECG and dry-electrode noise before the inverse transform; the diagonal layer supplies learned phase shifts, and the mixed activations let each layer choose between phase-preserving and symmetry-based nonlinearities.

What would settle it

Record a new dry-electrode abdominal dataset from pregnant women with simultaneous scalp-electrode fetal ECG reference, run CUNet on each channel, and compute PRD and PCC against the scalp reference; if the errors stay near the simulated level of about 26 percent PRD, the transfer claim holds, but a large increase would show the simulated results do not transfer. Alternatively, generate a test set with a different independently calibrated noise model; a sharp drop in PCC from the reported 95.9 would show the results depend on the specific noise model.

Watch

Extended reading notes

Core claim

The central claim is that preserving phase information in the time-frequency domain is the key to reliable single-channel fECG extraction. The proposed Complex U-Net applies a U-Net-style encoder-decoder to the complex short-time Fourier transform spectrogram of the abdominal recording, with complex convolutions, a learnable diagonal phase-shift layer, and a learnable convex combination of complex activation functions, then reconstructs the fetal signal by inverse short-time Fourier transform. The paper reports that, on its simulated dry-electrode test set, the method's percent RMS difference is less than half that of an attention U-Net and its correlation with the reference fetal ECG reaches about 96 percent, while R-peak detection F-score reaches 99.8 percent. On the wet-electrode benchmark dataset it achieves an F-score of 77.8 percent, comparable to the best baseline, and on four real dry-electrode volunteer recordings it produces visually recognizable fetal traces. The paper claims this combination, single-channel, dry textile electrode, complex-valued network, is new and moves toward non-invasive, self-administered fetal monitoring.

Load-bearing premise

The hand-built noise model, pink noise weighted 2, white noise 0.2, and a Gaussian mixture weighted 0.15, fitted to recordings from nine non-pregnant women, faithfully represents the noise and motion artefacts in abdominal recordings from pregnant women, and this model generates every simulated training and test recording.

Editorial extensions

If this is right

  • A single dry textile electrode can support clinically meaningful fECG extraction, not just heart-rate detection, because the network preserves P, QRS, and T morphology.
  • Real-valued deep networks leave substantial waveform distortion; processing the complex spectrogram is sufficient to reduce PRD by more than half and raise PCC to about 96 percent on simulated dry-electrode data.
  • The network generalizes to wet-electrode recordings without fine-tuning, suggesting the learned phase-aware representations transfer across electrode types and noise conditions.
  • The method maintains near-perfect R-peak detection on simulated data even at very low fetal-signal SNR, supporting fetal heart-rate estimation in early-gestation-like conditions.
  • A fully non-invasive, self-administered home monitoring system is plausibly within reach, pending large-scale clinical validation.

Reading between the lines

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

  • The strongest transfer test would be training the noise model on pregnant-woman dry-electrode recordings; if performance drops sharply, the simulated-to-real bridge is the fragile part of the claim.
  • The same complex-spectrogram architecture could be applied to other single-channel physiological mixtures, such as extracting fetal heart sounds or cleaning noisy wearable electrocardiograms.
  • A testable extension is to replace the hand-tuned noise weights with a learned noise generator or real noise augmentation, which could remove the model's dependence on empirical noise parameters.
  • Clinical interpretation, such as measuring QT intervals from extracted T-waves, is the real downstream test; the paper evaluates waveform reconstruction, not diagnostic accuracy.
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 / 4 minor

Summary. The paper proposes CUNet, a complex-valued UNet for extracting fetal ECG (fECG) from single-channel abdominal recordings acquired with dry textile electrodes. The authors create an in-silico training dataset by adding a parametric noise model, fitted to dry-electrode recordings from nine non-pregnant women, to synthetic maternal and fetal ECGs. They evaluate the method on 100 simulated test recordings generated from the same pipeline, on the PhysioNet 2013 wet-electrode benchmark, and on qualitative examples from four pregnant volunteers with dry electrodes. The reported in-silico results (PRD 26.2 ± 11.6, PCC 95.9 ± 5.7, F-score 99.8 ± 0.2) are used to support the claim of new state-of-the-art performance and accurate fECG morphology extraction across all evaluated settings.

Significance. If the central claim were established, the work would address a real clinical need: long-term, home-based fetal monitoring with comfortable dry textile electrodes and a single channel. The paper also contributes a synthetic data-generation pipeline and an external benchmark on PhysioNet, which is a useful evaluation practice. However, the evidence does not currently support the headline claims. The quantitative fECG-extraction results are obtained on simulated data generated by the same hand-tuned noise model used for training, the PhysioNet benchmark shows CUNet below the EKF baseline, and the real dry-electrode evaluation is purely visual with no reference signal. The complex-convolution formulation in Eq. (1) is also mathematically incorrect as written. These issues are load-bearing for the paper's main claims.

major comments (5)
  1. [Section II-A.1.b–c; Table II] The quantitative evidence supporting the headline claim is circular. The in-silico test set of 100 recordings is generated with the same parametric noise model (pink noise weight 2, white noise weight 0.2, Gaussian mixture weight 0.15; Section II-A.1.b) that was hand-tuned on nine non-pregnant women and used to generate the training data. The reported PRD 26.2 ± 11.6 and PCC 95.9 ± 5.7 therefore quantify how well CUNet inverts this particular noise generator, not how well it extracts fECG from real dry-electrode abdominal recordings in pregnancy. The real dry-electrode evaluation (Section III-B.2, Fig. 8) is purely qualitative, with no reference fECG, no detection metrics, and no comparison against baselines; the authors themselves state in Section IV-C that a large-scale clinical validation study is necessary. Consequently, the abstract's claims of "accurate extraction of fECG morphology across all evaluated settings" and of being the first to "effectively extract fECG signals from single-channel recordings using dry textile electrodes" are not supported by the evidence.
  2. [Section III-B.1 / Table III] The "new state-of-the-art results ... across all evaluated settings" claim is contradicted by the PhysioNet benchmark. On this dataset, EKF achieves F-score 80.0 ± 20.0 and SE 75.5 ± 23.8, while CUNet achieves F-score 77.8 ± 18.6 and SE 72.1 ± 22.5; the HRerr values (14.9 vs 15.4 bpm) overlap within their standard deviations. The discussion in Section IV-B.1 argues that EKF's higher HRerr "suggests it misses more fetal R-peaks compared to CUNet", but HRerr is an aggregate heart-rate error, not a missed-peak count, and the difference is negligible. The claim should be scaled back to "competitive with" rather than "superior to", or supported by statistical testing.
  3. [Section II-B.2, Eq. (1)] Equation (1) does not define a complex convolution. For a complex weight W = W^Re + iW^Im and complex input x, a genuine complex convolution has real part W^Re ∗ Re(x) − W^Im ∗ Im(x) and imaginary part W^Re ∗ Im(x) + W^Im ∗ Re(x). As written, Eq. (1) computes Y = W^Re ∗ Re(x) + i (W^Im ∗ Im(x)), which is equivalent to two independent real convolutions on the real and imaginary parts. The phase-coupling that the paper attributes to the convolutional layers therefore is not implemented by those layers; any coupling would have to come from the complex activations (Eqs. (3)–(6)) and the diagonal layer (Eq. (2)). The authors should either implement and report a true complex convolution or revise the text to describe the architecture as a two-branch real-valued network.
  4. [Section II-A.1.c / Section III-A / Table I] The SNR definitions are inconsistent. Dataset generation (Section II-A.1.c) states "SNR ranging from 5 to 20 dB", while Table I and Fig. 5 report SNR bins from −25 to 0 dB and the text explains that these values treat only the fECG as signal and mECG plus noise as noise. It is unclear whether the simulated recordings actually cover the low-SNR regime claimed in Table I, and whether the training data include such cases. This ambiguity affects the interpretation of the robustness claims.
  5. [Section III-B.2 / Fig. 8] The dry-electrode in-vivo evaluation is anecdotal. It shows two examples from four volunteers with no quantitative assessment: there is no reference fetal ECG, no beat-by-beat comparison, no test of heartbeat-count accuracy against an independent annotation, and no comparison with any baseline method on the same dry-electrode recordings. The dotted vertical lines in Fig. 8 appear to mark the authors' own detected beats, but the procedure for establishing those positions is not described. Without a quantitative or externally validated evaluation on dry-electrode data, the paper cannot support its central claim of accurate fECG extraction from dry textile electrodes.
minor comments (4)
  1. [Section II-D, Eq. (11)] The denominator in Eq. (11) is written as pP fECG(t)2 · qP ˆfECG(t)2; the square-root symbols are missing, so the formula should read sqrt(Σ fECG(t)^2) · sqrt(Σ ˆfECG(t)^2).
  2. [Section III-B.2 / Fig. 8 captions] The caption text "Aquired Signal" should be "Acquired Signal"; additionally, the main text does not state how many recordings from the four volunteers are shown or how the displayed segments were selected.
  3. [Section II-A.1.a] The text refers to the physiological frequency band as 0–100 Hz but then describes a 1–100 Hz bandpass filter; this small discrepancy should be reconciled.
  4. [Section I / Section II-C] The comparison set omits recently published single-channel deep-learning methods that are cited in the introduction, such as TCGAN [18] and W-NETR [21]; if the goal is to claim state-of-the-art performance, these methods should be included in the benchmark or a justification for their exclusion should be given.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the simulation-based evaluation is self-contained, the PhysioNet benchmark provides independent evidence, and the main threats are external validity and an implementation issue, not circular derivation.

full rationale

The paper's central derivation is not circular. The in-silico test set (Section II-A.1.c) is drawn from the same author-defined noise model used for training, and the noise weights (pink=2, white=0.2, Gaussian-mixture=0.15) are 'selected based on empirical observations' (Section II-A.1.b). This is a genuine external-validity limitation: the reported PRD 26.2 ± 11.6 and PCC 95.9 ± 5.7 (Table II) measure performance on the authors' synthetic distribution, not on real dry-electrode abdominal recordings from pregnant women, and the real dry-electrode evaluation (Section III-B.2, Fig. 8) is qualitative only. However, this is not circularity by the paper's own equations: the test metrics are not equivalent to the noise-model parameters, the network could have scored arbitrarily worse, and the same synthetic test set is used for all compared methods. The PhysioNet benchmark (wet electrodes, independent annotations) gives an external check, although CUNet (F-score 77.8 ± 18.6) is slightly below EKF (80.0 ± 20.0) in Table III, which undercuts the abstract's 'state-of-the-art across all evaluated settings' but is not a circular step. Section IV-C explicitly concedes that 'a large-scale clinical validation study would be necessary before employing it in clinical settings.' Self-citations to [12]-[14] are used as tools (noise removal, SNR ranges) and are not load-bearing circularity. Separately, Eq. (1) omits the cross terms of a true complex convolution, so the stated layer is equivalent to two independent real convolutions on real and imaginary parts; this is a technical correctness flaw in the claimed phase-aware mechanism, not a derivation that reduces a prediction to an input. Overall, no specific circular step was identified.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central quantitative result depends on a synthetic dataset whose noise model is assembled by hand from the authors' own dry-electrode recordings. The fecgsyn generator supplies ECG morphology, but no independent evidence shows the simulated signals reproduce real fetal ECG morphology or real pregnancy noise. The PhysioNet dataset provides only a wet-electrode external check, and it does not support the dry-electrode claim.

free parameters (3)
  • noise component weights = pink=2, white=0.2, Gaussian mixture=0.15
    Hand-chosen based on empirical observations of real dry-electrode noise; these weights shape the simulation used for both training and evaluation.
  • Gaussian mixture noise parameters = sigma0=1, sigma1=10, p=0.1
    Chosen by hand to model background and impulse noise; no fitting procedure is described.
  • random noise band ranges = pink 0 to 9-12 Hz; white 10 to 60-90 Hz
    Randomized within manually selected intervals to increase variability, not derived from a validated model.
assumptions (4)
  • domain assumption Synthetic fecgsyn signals faithfully represent real fetal and maternal ECG morphology
    The simulator is used as ground truth for training and quantitative evaluation; no comparison to real fECG morphology is provided.
  • domain assumption Noise extracted from non-pregnant women after BSS-based ECG removal represents noise during pregnancy with dry electrodes
    Section II-A.1.a uses recordings from nine non-pregnant women; the noise is assumed to transfer to pregnant abdominal recordings.
  • domain assumption Noise is additive and decomposable into independent pink, white, and Gaussian mixture components
    Section II-A.1.b simulates noise as a weighted sum of normalized spectral components; this model generates both training and test sets.
  • ad hoc to paper A separate-channel real/imaginary convolution is a valid complex convolution
    Eq. (1) defines output using only W_Re * Re(x) and W_Im * Im(x), omitting cross terms of a standard complex convolution; the claim that the network processes phase relationships rests on this nonstandard definition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Complex UNet Approach for Non-Invasive Fetal ECG Extraction Using Single-Channel Dry Textile Electrodes." pith.science (2026). https://pith.science/paper/PNYZCLPL

@misc{pith2026250622457,
  author       = {Pith},
  title        = {Pith review of: A Complex UNet Approach for Non-Invasive Fetal ECG Extraction Using Single-Channel Dry Textile Electrodes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PNYZCLPL}},
  note         = {Machine review of arXiv:2506.22457}
}
read the original abstract

Continuous, non-invasive pregnancy monitoring is crucial for minimising potential complications. The fetal electrocardiogram (fECG) represents a promising tool for assessing fetal health beyond clinical environments. Home-based monitoring necessitates the use of a minimal number of comfortable and durable electrodes, such as dry textile electrodes. However, this setup presents many challenges, including increased noise and motion artefacts, which complicate the accurate extraction of fECG signals. To overcome these challenges, we introduce a pioneering method for extracting fECG from single-channel recordings obtained using dry textile electrodes using AI techniques. We created a new dataset by simulating abdominal recordings, including noise closely resembling real-world characteristics of in-vivo recordings through dry textile electrodes, alongside mECG and fECG. To ensure the reliability of the extracted fECG, we propose an innovative pipeline based on a complex-valued denoising network, Complex UNet. Unlike previous approaches that focused solely on signal magnitude, our method processes both real and imaginary components of the spectrogram, addressing phase information and preventing incongruous predictions. We evaluated our novel pipeline against traditional, well-established approaches, on both simulated and real data in terms of fECG extraction and R-peak detection. The results showcase that our suggested method achieves new state-of-the-art results, enabling an accurate extraction of fECG morphology across all evaluated settings. This method is the first to effectively extract fECG signals from single-channel recordings using dry textile electrodes, making a significant advancement towards a fully non-invasive and self-administered fECG extraction solution.

Figures

Figures reproduced from arXiv: 2506.22457 by the authors.

Figure 1
Figure 1. Illustrative case: generated noise signal in frequency domain intended for replication. c) Dataset generation: the synthetic dataset was obtained by superimposing 10.100 clean one-minute synthetic mECG and fECG signals as produced through fecgsyn generator [22] together with additive model-based noise, independently generated for each recording. For each acquisition, we select a random value from specific intervals,… view at source ↗
Figure 2
Figure 2. The dataset creation pipeline. We start by extracting the noise from real dry ECG recordings using signal processing and Blind Source Separation methods. After determining the noise characteristics, we sample noise alongside simulated maternal and fetal ECG to create our simulated dataset. followed by a general description of the used architecture. Further sections are dedicated to a more in-depth analysis of the ne… view at source ↗
Figure 3
Figure 3. Our approach: First, we apply a time-frequency analysis on sECG to obtain its spectrogram SECG. We propagate it through the ComplexUNet, denoted by H, and obtain a denoised fetal extraction SfECG = H(SECG). The last step is composed of an inverse STFT, resulting in the fECG signal. Note that each layer of our networks works in the Complex domain. Another important aspect is represented by the presence of skip connec… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Examples of abdominal simulated (noisy, in green) signals alongside the reference (blue) and predicted fECGs (red), at different PRD levels: (a) low, (b) medium, (c) high. All signals are normalised for a better visualisation. process is more challenging. Despite the n…
Figure 5
Figure 5. Figure 5: Boxplots of PRD (blue) and PCC (orange) values w.r.t. the SNR for CUNet TABLE III EVALUATION METRICS FOR R-PEAK DETECTION (F-SCORE, SE, AND HRERR ) FOR THE PROPOSED METHOD COMPARED TO BENCHMARK ALGORITHMS. ALL VALUES REPRESENT SINGLE-CHANNEL ESTIMATES COMPUTED ON THE P…
Figure 6
Figure 6. Figure 6: Visual comparison between the reference (blue) and predicted fECG (red), using different neural-based methods. Top represents CUNet, middle is AttUNet, and bottom is UNet. The superior performance of the proposed CUNet, com￾pared to the more basic UNet and AttUNet, is …
Figure 4
Figure 4. Figure 4: when the SNR is high (see panel (a)) - typically ar the [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 7
Figure 7. Figure 7: Visual inspection of the results obtained using UNet (blue), AttUNet (red) and CUNet (green) against different benchmarking algo￾rithms (black) on the PhysioNet dataset. real-value representations, they do not capture phase-related features that are fundamental for fEC…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 45 canonical work pages

  1. [1]

    A neglected tragedy the global burden of stillbirths: report of the UN inter-agency group for child mortality estimation,

    L. Hug, A. Mishra, S. Lee, D. You, A. Moran, K. L. Strong, and B. Cao, “A neglected tragedy the global burden of stillbirths: report of the UN inter-agency group for child mortality estimation,” 2020

    work page 2020

  2. [2]

    A review of fetal cardiac monitoring, with a focus on low- and middle-income countries,

    C. E. Valderrama, N. Ketabi, F. Marzbanrad, P. Rohloff, and G. D. Clifford, “A review of fetal cardiac monitoring, with a focus on low- and middle-income countries,” Physio. measurement, vol. 41, no. 11, p. 11TR01, 2020

    work page 2020

  3. [3]

    Maternal-fetal conditions necessi- tating a medical intervention resulting in preterm birth,

    C. V . Ananth and A. M. Vintzileos, “Maternal-fetal conditions necessi- tating a medical intervention resulting in preterm birth,” American J. of Obstetrics and Gynecology , vol. 195, no. 6, pp. 1557–1563, 2006

    work page 2006

  4. [4]

    Continuous cardiotocography (CTG) as a form of electronic fetal monitoring (EFM) for fetal assessment during labour,

    Z. Alfirevic, G. M. Gyte, A. Cuthbert, and D. Devane, “Continuous cardiotocography (CTG) as a form of electronic fetal monitoring (EFM) for fetal assessment during labour,” Physio. measurement, vol. 2, no. 2, 2017

    work page 2017

  5. [5]

    Fetal heart rate monitoring implemented by dynamic adaptation of transmission power of a flexible ultrasound transducer array,

    P. Hamelmann, M. Mischi, A. F. Kolen, J. O. Van Laar, R. Vullings, and J. W. Bergmans, “Fetal heart rate monitoring implemented by dynamic adaptation of transmission power of a flexible ultrasound transducer array,” Sensors, vol. 19, no. 5, p. 1195, 2019

    work page 2019

  6. [6]

    Clin- ically accurate fetal ecg parameters acquired from maternal abdominal sensors,

    G. Clifford, R. Sameni, J. Ward, J. Robinson, and A. J. Wolfberg, “Clin- ically accurate fetal ecg parameters acquired from maternal abdominal sensors,” American J. of Obstetrics and Gynecology , vol. 205, no. 1, pp. 47–51, 2011

    work page 2011

  7. [7]

    A review of signal processing techniques for non-invasive fetal electrocardiography,

    R. Kahankova, R. Martinek, R. Jaros, K. Behbehani, A. Matonia, M. Jezewski, and J. A. Behar, “A review of signal processing techniques for non-invasive fetal electrocardiography,” IEEE Reviews in Biomed. Eng., vol. 13, pp. 51–73, 2019

    work page 2019

  8. [8]

    A systematic review of cardiac time intervals utilising non-invasive fetal electrocardiogram in normal fetuses,

    V . Smith, S. Arunthavanathan, A. Nair, D. Ansermet, F. da Silva Costa, and E. M. Wallace, “A systematic review of cardiac time intervals utilising non-invasive fetal electrocardiogram in normal fetuses,” BMC Pregnancy and Childbirth, vol. 18, pp. 1–15, 2018

    work page 2018

Show all 46 references
  1. [9]

    Unsupervised denoising of the non-invasive fetal electrocardiogram with sparse domain Kalman filtering and vectorcardiographic loop alignment,

    I. de Vries, J. O. van Laar, R. Vullings et al., “Unsupervised denoising of the non-invasive fetal electrocardiogram with sparse domain Kalman filtering and vectorcardiographic loop alignment,” Physiological Mea- surement, vol. 45, no. 7, p. 075006, 2024

    2024

  2. [10]

    fECG extraction using 1D convolution neural network,

    Y . Sharma, S. Ray, and O. P. Yadav, “fECG extraction using 1D convolution neural network,” Intelligent Computing and Communication Systems, pp. 331–338, 2021

    2021

  3. [11]

    A deep learning framework for noninvasive fetal ECG signal extraction,

    M. Wahbah, M. S. Zitouni, R. Al Sakaji, K. Funamoto, N. Widatalla, A. Krishnan, Y . Kimura, and A. H. Khandoker, “A deep learning framework for noninvasive fetal ECG signal extraction,” Frontiers in Physiology, vol. 15, p. 1329313, 2024

    2024

  4. [12]

    Dedicated algorithm for unobtrusive fetal heart rate monitoring using multiple dry electrodes,

    A. Galli, E. Peri, Y . Zhang, R. Vullings, M. van der Ven, G. Giorgi, S. Ouzounov, P. J. Harpe, and M. Mischi, “Dedicated algorithm for unobtrusive fetal heart rate monitoring using multiple dry electrodes,” Sensors, vol. 21, no. 13, p. 4298, 2021

    2021

  5. [13]

    Automatic optimization of multichannel electrode configurations for robust fetal heart rate detection by blind source separation,

    A. Galli, E. Peri, C. Rabotti, S. Ouzounov, and M. Mischi, “Automatic optimization of multichannel electrode configurations for robust fetal heart rate detection by blind source separation,” IEEE Trans. on Biomed. Eng., vol. 70, no. 4, pp. 1196–1207, 2022

    2022

  6. [14]

    Improved mECG removal and fECG extraction by integrated periodic components analysis and singular value decomposition,

    A. Galli, E. Peri, P. Hamelmann, and M. Mischi, “Improved mECG removal and fECG extraction by integrated periodic components analysis and singular value decomposition,” in IEEE Int. Symposium on Medical Measurements and Applications , 2024, pp. 1–6

    2024

  7. [15]

    A deep learning approach for fetal QRS complex detection,

    W. Zhong, L. Liao, X. Guo, and G. Wang, “A deep learning approach for fetal QRS complex detection,” Physiological Measurement, vol. 39, no. 4, p. 045004, 2018

    2018

  8. [16]

    Fetal QRS detection based on convolutional neural networks in noninvasive fetal electrocar- diogram,

    J. S. Lee, M. Seo, S. W. Kim, and M. Choi, “Fetal QRS detection based on convolutional neural networks in noninvasive fetal electrocar- diogram,” in Proc. of Int. Conf. on Frontiers of Signal Process. , 2018, pp. 75–78

    2018

  9. [17]

    Using U-Nets for Accurate R-Peak Detection in Fetal ECG Recordings,

    P. Zhou, S. So, and B. Schwerin, “Using U-Nets for Accurate R-Peak Detection in Fetal ECG Recordings,” vol. 14, no. 09, pp. 1–10, 2024

    2024

  10. [18]

    TCGAN: Temporal Convolutional Generative Adversarial Network for Fetal ECG Extraction Using Single-Channel Abdominal ECG,

    Z.-Z. Huang, W.-T. Zhang, Y . Li, J. Cui, and Y .-R. Zhang, “TCGAN: Temporal Convolutional Generative Adversarial Network for Fetal ECG Extraction Using Single-Channel Abdominal ECG,” IEEE J. of Biomed. and Health Info. , vol. 29, no. 5, pp. 3478–3487, 2025

    2025

  11. [19]

    A complete adaptive method for fetal ECG extraction based on single channel,

    M. Zhang and G. Wei, “A complete adaptive method for fetal ECG extraction based on single channel,” in J. of Physics: Conference Series, vol. 1621, no. 1, 2020, pp. 12–19

    2020

  12. [20]

    Fetal ECG extraction by extended state kalman filtering based on single-channel recordings,

    M. Niknazar, B. Rivet, and C. Jutten, “Fetal ECG extraction by extended state kalman filtering based on single-channel recordings,” IEEE Trans. on Biomed. Eng. , vol. 60, no. 5, pp. 1345–1352, 2012

    2012

  13. [21]

    One-dimensional W-NETR for non-invasive single channel fetal ECG extraction,

    M. Almadani, L. Hadjileontiadis, and A. Khandoker, “One-dimensional W-NETR for non-invasive single channel fetal ECG extraction,” IEEE J. of Biomed. and Health Info. , vol. 27, no. 7, pp. 3198–3209, 2023

    2023

  14. [22]

    An open-source framework for stress-testing non-invasive foetal ECG extraction algorithms,

    F. Andreotti, J. Behar, S. Zaunseder, J. Oster, and G. D. Clifford, “An open-source framework for stress-testing non-invasive foetal ECG extraction algorithms,” Physiological Measurement , vol. 37, no. 5, p. 627, 2016

    2016

  15. [23]

    Fetal electrocardiograms, direct and abdominal with reference heartbeat annotations,

    A. Matonia, J. Jezewski, T. Kupka, M. Jezewski, K. Horoba, J. Wrobel, R. Czabanski, and R. Kahankowa, “Fetal electrocardiograms, direct and abdominal with reference heartbeat annotations,” Nature Scientific Data, vol. 7, no. 1, p. 200, 2020

    2020

  16. [24]

    A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, “PhysioBank, PhysioToolkit, and PhysioNet: components of a 12 preprint submitted to IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS ...

    2000

  17. [25]

    Noninvasive fetal ECG: the PhysioNet/computing in cardiology challenge ,

    I. Silva, J. Behar, R. Sameni, T. Zhu, J. Oster, G. D. Clifford, and G. B. Moody, “Noninvasive fetal ECG: the PhysioNet/computing in cardiology challenge ,” in IEEE Computing in Cardiology , 2013, pp. 149–152

    2013

  18. [26]

    Complex-Valued Neural Networks: A Comprehensive Survey,

    C. Lee, H. Hasegawa, and S. Gao, “Complex-Valued Neural Networks: A Comprehensive Survey,” J. of Automatica Sinica , vol. 9, no. 8, pp. 1406–1426, 2022

    2022

  19. [27]

    Theory and implementation of complex-valued neural networks,

    J. A. Barrachina, C. Ren, G. Vieillard, C. Morisseau, and J.-P. Ovar- lez, “Theory and implementation of complex-valued neural networks,” preprint arXiv:2302.08286, 2023

    2023 arXiv

  20. [28]

    Complex-valued Neurons Can Learn More but Slower than Real-valued Neurons via Gradient Descent,

    J.-H. Wu, S.-Q. Zhang, Y . Jiang, and Z.-H. Zhou, “Complex-valued Neurons Can Learn More but Slower than Real-valued Neurons via Gradient Descent,” in Proc. Adv. in Neural Inf. Process. Syst. , vol. 36, 2023, pp. 23 714–23 747

    2023

  21. [29]

    Appli- cations of Neural Networks in Biomedical Data Analysis,

    R. Weiss, S. Karimijafarbigloo, D. Roggenbuck, and S. R ¨odiger, “Appli- cations of Neural Networks in Biomedical Data Analysis,”Biomedicines, vol. 10, no. 7, p. 1469, 2022

    2022

  22. [30]

    Complex-Valued Vs. Real-Valued Neural Networks for Classification Perspectives: An Example on Non-Circular Data,

    J. A. Barrachina, C. Ren, C. Morisseau, G. Vieillard, and J.-P. Ovarlez, “Complex-Valued Vs. Real-Valued Neural Networks for Classification Perspectives: An Example on Non-Circular Data,” in Proc. Int. Conf. Acoust. Speech Signal Process. , 2021, pp. 2990–2994

    2021

  23. [31]

    Analysis of deep complex-valued convolutional neural networks for MRI reconstruction and phase-focused applications,

    E. Cole, J. Cheng, J. Pauly, and S. Vasanawala, “Analysis of deep complex-valued convolutional neural networks for MRI reconstruction and phase-focused applications,” Magnetic Resonance in Medicine , vol. 86, no. 2, pp. 1093–1109, 2021

    2021

  24. [32]

    A survey on fECG extraction using neural network and adaptive filter,

    A. Mohammed Kaleem and R. D. Kokate, “A survey on fECG extraction using neural network and adaptive filter,” Soft Computing , vol. 25, pp. 4379–4392, 2021

    2021

  25. [33]

    U-Net: Convolutional Net- works for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Net- works for Biomedical Image Segmentation,” in Med. Image Computing and Computer-Assisted Intervention , vol. 18, 2015, pp. 234–241

    2015

  26. [34]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in Proc. IEEE Conf. Comput. Vis. Pattern Recogn., 2016, pp. 770–778

    2016

  27. [35]

    Design of Robust Complex-Valued Feed-Forward Neural Networks,

    A. Neacsu, R. Ciubotaru, J.-C. Pesquet, and C. Burileanu, “Design of Robust Complex-Valued Feed-Forward Neural Networks,” in Proc. European. Signal. Process. Conf. , 2022, pp. 1596–1600

    2022

  28. [36]

    Connecting Pre-trained Language Model and Downstream Task via Properties of Representation,

    C. Wu, H. Lee, and R. Ge, “Connecting Pre-trained Language Model and Downstream Task via Properties of Representation,” in Proc. Adv. in Neural Inf. Process. Syst. , 2023, pp. 2044–2067

    2023

  29. [37]

    Complex domain backpropagation,

    G. Georgiou and C. Koutsougeras, “Complex domain backpropagation,” IEEE Trans. on Circuits and Systems II , vol. 39, no. 5, pp. 330–334, 1992

    1992

  30. [38]

    A novel ECG denoising scheme using the ensemble kalman filter,

    S. Sarafan, H. Vuong, D. Jilani, S. Malhotra, M. P. Lau, M. Vishwanath, T. Ghirmai, and H. Cao, “A novel ECG denoising scheme using the ensemble kalman filter,” Proc. IEEE Int. Conf. of Eng. in Med. & Bio. Soc., pp. 2005–2008, 2022

    2005

  31. [39]

    A dynamical model for generating synthetic electrocardiogram signals,

    P. E. McSharry, G. D. Clifford, L. Tarassenko, and L. A. Smith, “A dynamical model for generating synthetic electrocardiogram signals,” IEEE Trans. on Biomed. Eng. , vol. 50, no. 3, pp. 289–294, 2003

    2003

  32. [40]

    EEG artifact removal—state- of-the-art and guidelines,

    J. A. Urig ¨uen and B. Garcia-Zapirain, “EEG artifact removal—state- of-the-art and guidelines,” J. of Neur. Eng. , vol. 12, no. 3, p. 031001, 2015

    2015

  33. [41]

    An efficient unsupervised fetal QRS complex detection from abdominal maternal ECG,

    M. Varanini, G. Tartarisco, L. Billeci, A. Macerata, G. Pioggia, and R. Balocchi, “An efficient unsupervised fetal QRS complex detection from abdominal maternal ECG,” Physiological measurement , vol. 35, no. 8, p. 1607, 2014

    2014

  34. [42]

    Detection of fetal arrhythmia by adaptive single channel electrocardiogram extraction,

    M. Suganthy, S. I. Joy, and P. Anandan, “Detection of fetal arrhythmia by adaptive single channel electrocardiogram extraction,” J. Physical and Eng. Sci. in Med. , vol. 44, no. 3, pp. 683–692, 2021

    2021

  35. [43]

    RPnet: A Deep Learning approach for robust R Peak detection in noisy ECG,

    S. Vijayarangan, R. Vignesh, B. Murugesan, S. Preejith, J. Joseph, and M. Sivaprakasam, “RPnet: A Deep Learning approach for robust R Peak detection in noisy ECG,” in Proc. Int. Conf. IEEE Eng. in Medicine , 2020, pp. 345–348

    2020

  36. [44]

    Fetal ECG extraction from maternal ECG using deeply supervised LinkNet++ model,

    A. Rahman, S. Mahmud, M. E. Chowdhury, H. C. Yalcin, A. Khandakar, O. Mutlu, Z. B. Mahbub, R. Y . Kamal, and S. Pedersen, “Fetal ECG extraction from maternal ECG using deeply supervised LinkNet++ model,” Eng. Applications of Artificial Intell., vol. 123, p. 106414, 2023

    2023

  37. [45]

    Attention U-Net: Learning Where to Look for the Pancreas,

    O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y . Hammerla, B. Kainz, B. Glocker, and D. Rueckert, “Attention U-Net: Learning Where to Look for the Pancreas,” in Proc. Medical Imaging with Deep Learning , 2018, pp. 1–10

    2018

  38. [46]

    Amer- ican national standard for ambulatory electrocardiographs, publication ansi,

    A. ssociation for the Advancement of Medical Instrumentation, “Amer- ican national standard for ambulatory electrocardiographs, publication ansi,” AAMI EC38-1994, 1994

    1994

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.