The reviewed record of science sign in
Pith

arxiv: 2605.29695 · v1 · pith:XYJJOR5A · submitted 2026-05-28 · cs.AI · cs.CE· cs.LG· math.PR

FHRFormer: A Self-Supervised Masked Transformer Framework for Fetal Heart Rate Time-Series Inpainting and Forecasting

Kjersti Engan , Neel Kanwal , Anita Yeconia , Ladislaus Blacy , Yuda Munyaw , Estomih Mduma , Hege Ersdal This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-29 07:02 UTCgrok-4.3pith:XYJJOR5Arecord.jsonopen to challenge →

classification cs.AI cs.CEcs.LGmath.PR
keywords fetal heart ratemasked transformerself-supervised learningsignal inpaintingtime-series forecastingautoencodermissing data reconstructionwearable monitoring
0
0 comments X

The pith

A masked transformer autoencoder reconstructs missing fetal heart rate signals by capturing temporal and frequency components.

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

The paper introduces a self-supervised masked transformer autoencoder to fill gaps in fetal heart rate recordings that arise from sensor displacement in wearable monitors. Simple interpolation often distorts the signal's frequency content, limiting reliable AI analysis of large FHR datasets for predicting birth risks such as the need for breathing assistance. The method learns to reconstruct missing segments by attending to both local time patterns and frequency information. It handles gaps of varying lengths and supports both inpainting and forward prediction. Successful application would allow cleaner retrospective datasets for risk models and eventual on-device use in continuous monitoring.

Core claim

The authors establish that a masked transformer-based autoencoder trained self-supervised on FHR data reconstructs missing signals by capturing both local temporal and frequency components, demonstrating robustness across different durations of missing data for inpainting and forecasting tasks.

What carries the argument

Masked transformer-based autoencoder that masks portions of the input time series and learns to reconstruct them while modeling temporal sequences and frequency content.

If this is right

  • The approach can be applied retrospectively to research datasets to support development of AI-based algorithms for predicting risk of needing breathing assistance at birth.
  • It addresses the failure of interpolation methods to preserve spectral characteristics of FHR signals.
  • Future integration into wearable FHR monitors could enable earlier and more robust risk detection during labor.
  • It permits continuous fetal monitoring without data loss from maternal movement or position changes.

Where Pith is reading between the lines

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

  • The same reconstruction approach could apply to other intermittently dropout-prone biomedical signals such as ECG or blood pressure traces.
  • Real-time device integration would require separate checks on computational latency and power use under live conditions.
  • Combining inpainted signals directly with labeled outcome data might improve end-to-end models for obstetric intervention prediction.

Load-bearing premise

That a standard masked transformer autoencoder trained in a self-supervised manner on FHR data will successfully preserve clinically relevant spectral and temporal features without additional domain-specific constraints or labeled examples.

What would settle it

A test set experiment in which the power spectrum or variability metrics of the reconstructed signals differ significantly from those of complete original recordings across multiple gap lengths would falsify the robustness claim.

Figures

Figures reproduced from arXiv: 2605.29695 by Anita Yeconia, Estomih Mduma, Hege Ersdal, Kjersti Engan, Ladislaus Blacy, Neel Kanwal, Yuda Munyaw.

Figure 1
Figure 1. Figure 1: The application of transformer-based autoencoders in this research work: The architecture is trained on the preprocessed version of FHR data, which can perform inpainting and forecasting tasks. Data preparation: The pooled datasets are preprocessed for Doppler noise and linearly interpolated before performing min-max normalization. Reconstruction: A transformer-based encoder-decoder trained in a self-super… view at source ↗
Figure 2
Figure 2. Figure 2: An overview of the ’FHRFormer’ architecture. The architecture takes the fetal heart rate signal and divides it into patches. The patches are then embedded and masked before linearly projecting into the encoder. The encoder has five transformer blocks, which produce a compact representation that is then used by the decoder, which contains five transformer blocks. The final reconstruction is compared with th… view at source ↗
Figure 3
Figure 3. Figure 3: Model performance trends across varying masking ratios (x-axis) with fixed input size 30. The six subplots display Reconstruction Loss (RL), Peak Signal-to-Noise Ratio (PSNR), SSIM, Frechet Distance (FID), Root Mean Squared Error (RMSE), and Correlation Coefficient over y-axes. Each subplot illustrates how the metric changes with masking ratio (γ), highlighting the trade-offs in model accuracy and quality.… view at source ↗
Figure 4
Figure 4. Figure 4: Inpainting application using the FHRFormer. The first row shows the original FHR signal from the Moyo device. Vertical black lines indicate samples removed during preprocessing (denoising process). The second row shows the noise-removed and the interpolated version. The third and fourth rows show inpainting performance by the best (Hybrid-30) and the second-best (Hybrid-60). 3.1 Self-supervised FHR Encoder… view at source ↗
Figure 5
Figure 5. Figure 5: Forecasting application using the FHRFormer. The FHRFormer is fed with a context window (gray region), which is 3600 timesteps (30 minutes of FHR data) as the past horizon, and starts forecasting 30 timesteps (15 seconds of FHR data) in a progressive forecasting style. The first row shows forecasting performance by the TimeGPT (Garza et al., 2023) model. The second row shows forecasting performance by the … view at source ↗
read the original abstract

Approximately 10% of newborns require assistance to initiate breathing at birth, and around 5% need ventilation support. Fetal heart rate (FHR) monitoring plays a crucial role in assessing fetal well-being during prenatal care, enabling the detection of abnormal patterns and supporting timely obstetric interventions to mitigate fetal risks during labor. Applying artificial intelligence (AI) methods to analyze large datasets of continuous FHR monitoring episodes with diverse outcomes may offer novel insights into predicting the risk of needing breathing assistance or interventions. Recent advances in wearable FHR monitors have enabled continuous fetal monitoring without compromising maternal mobility. However, sensor displacement during maternal movement, as well as changes in fetal or maternal position, often lead to signal dropout, resulting in gaps in recorded FHR data. Such missing data limits the extraction of meaningful insights and complicates automated (AI-based) analysis. Traditional approaches to handling missing data, such as simple interpolation techniques, often fail to preserve the spectral characteristics of the signals. In this paper, we propose a masked transformer-based autoencoder approach to reconstruct missing FHR signals by capturing both local temporal and frequency components of the data. The proposed method demonstrates robustness across varying durations of missing data and can be used for signal inpainting and forecasting. The proposed approach can be applied retrospectively to research datasets to support the development of AI-based risk algorithms. In the future, the proposed method could be integrated into wearable FHR monitoring devices to achieve earlier and more robust risk detection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper proposes FHRFormer, a self-supervised masked transformer-based autoencoder for inpainting and forecasting fetal heart rate (FHR) time-series signals. It addresses gaps from sensor dropout in wearable monitors by capturing local temporal and frequency components, claiming robustness across varying missing-data durations for retrospective use in AI-based fetal risk algorithms and potential future device integration.

Significance. If the empirical claims hold, the work could meaningfully improve handling of incomplete FHR recordings compared with interpolation, supporting downstream AI models for neonatal intervention prediction. The self-supervised formulation is well-matched to the unlabeled nature of continuous monitoring data.

major comments (1)
  1. [Abstract] Abstract: the central claim that the method 'demonstrates robustness across varying durations of missing data' is unsupported by any quantitative results, baselines, error metrics, dataset descriptions, or experimental sections in the provided manuscript text, rendering the primary contribution unevaluable.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'capturing both local temporal and frequency components' is stated without indicating the architectural mechanism (e.g., explicit spectral layers, Fourier features, or learned filters).

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their review and for highlighting this important issue with the abstract. We address the comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the method 'demonstrates robustness across varying durations of missing data' is unsupported by any quantitative results, baselines, error metrics, dataset descriptions, or experimental sections in the provided manuscript text, rendering the primary contribution unevaluable.

    Authors: We agree that the abstract claim is unsupported by any quantitative evidence, baselines, metrics, datasets, or experimental sections in the manuscript text. This renders the primary contribution unevaluable from the provided material. We will revise the manuscript by either removing or qualifying the unsupported claim in the abstract, or by adding the required experimental results, baselines, error metrics, and dataset descriptions to substantiate it. The revised version will ensure the contribution is properly supported and evaluable. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes a standard self-supervised masked transformer autoencoder for FHR time-series inpainting and forecasting. No equations, parameter-fitting procedures, derivations, or self-citations appear in the abstract or method outline that reduce any claimed prediction or result to its own inputs by construction. The approach relies on established transformer masking techniques applied to the domain, without any load-bearing step that renames a fit as a prediction or imports uniqueness via prior author work. The derivation chain is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no information on free parameters, axioms, or invented entities; ledger is therefore empty.

pith-pipeline@v0.9.1-grok · 5834 in / 1159 out tokens · 34018 ms · 2026-06-29T07:02:20.751863+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

43 extracted references · 4 canonical work pages

  1. [1]

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

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

  2. [2]

    write newline

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

  3. [3]

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

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key language month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid...

  4. [4]

    write newline

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

  5. [5]

    Alfirevic, Z., Gyte, G. M. L., Cuthbert, A., and Devane, D. (2017). Continuous cardiotocography ( CTG ) as a form of electronic fetal monitoring ( EFM ) for fetal assessment during labour. Cochrane Database of Systematic Reviews 2, CD006066. doi:10.1002/14651858.CD006066.pub3 alfirevic2017

    work page doi:10.1002/14651858.cd006066.pub3 2017

  6. [6]

    and Imtiaz, M

    Alim, A. and Imtiaz, M. H. (2023). Wearable sensors for the monitoring of maternal health—a systematic review. Sensors 23, 2411 alim2023wearable

    2023

  7. [7]

    Barzideh, F., Urdal, J., Hussein, K., Engan, K., Skretting, K., Mdoe, P., et al. (2018). Estimation of missing data in fetal heart rate signals using shift-invariant dictionary. In 2018 26th European Signal Processing Conference (EUSIPCO) (IEEE), 762--766 barzideh2018estimation

    2018

  8. [8]

    Boudet, S., Houz \'e de l’Aulnoit, A., Peyrodie, L., Demailly, R., and Houz \'e de l’Aulnoit, D. (2022). Use of deep learning to detect the maternal heart rate and false signals on fetal heart rate recordings. Biosensors 12, 691 boudet2022use

    2022

  9. [9]

    Campos, I., Gon c alves, H., Bernardes, J., and Castro, L. (2024). Fetal heart rate preprocessing techniques: A scoping review. Bioengineering 11, 368 campos2024fetal

    2024

  10. [10]

    Cesarelli, M., Romano, M., Ruffo, M., Bifulco, P., Pasquariello, G., and Fratini, A. (2011). Psd modifications of fhrv due to interpolation and ctg storage rate. Biomedical Signal Processing and Control 6, 225--230 cesarelli2011psd

    2011

  11. [11]

    Cockburn, N., Osborne, C., Withana, S., Elsmore, A., Nanjappa, R., South, M., et al. (2024). Clinical decision support systems for maternity care: a systematic review and meta-analysis. EClinicalMedicine 76 cockburn2024clinical

    2024

  12. [12]

    Da Poian, G., Bernardini, R., and Rinaldo, R. (2015). Sparse representation for fetal qrs detection in abdominal ecg recordings. In 2015 E-Health and Bioengineering Conference (EHB) (IEEE), 1--4 da2015sparse

    2015

  13. [13]

    G., and Djuri \'c , P

    Dash, S., Quirk, J. G., and Djuri \'c , P. M. (2014). Fetal heart rate classification using generative models. IEEE Transactions on Biomedical Engineering 61, 2796--2805 dash2014fetal

    2014

  14. [14]

    Dawes, A., Ameh, C., Moxon, S., Gannon, H., Jha, A., Basnet, O., et al. (2021). Neonatal resuscitation and immediate newborn assessment and stimulation for the prevention of neonatal deaths: A systematic review, meta-analysis and delphi estimation of mortality effect. BMC Public Health 21, 1--14. doi:10.1186/s12889-020-10033-8. Published online December 2...

    work page doi:10.1186/s12889-020-10033-8 2021

  15. [15]

    Garza, A., Challu, C., and Mergenthaler-Canseco, M. (2023). Timegpt-1. arXiv preprint arXiv:2310.03589 garza2023timegpt

    work page Pith review arXiv 2023

  16. [16]

    J., Moulden, M., and Redman, C

    Georgieva, A., Payne, S. J., Moulden, M., and Redman, C. W. (2013). Artificial neural networks applied to fetal monitoring in labour. Neural Computing and Applications 22, 85--93 georgieva2013artificial

    2013

  17. [17]

    K., Catelli, D

    Ghosh, A. K., Catelli, D. S., Wilson, S., Nowlan, N. C., and Vaidyanathan, R. (2024). Multi-modal detection of fetal movements using a wearable monitor. Information Fusion 103, 102124 ghosh2024multi

    2024

  18. [18]

    Jiang, L., Dai, B., Wu, W., and Loy, C. C. (2021). Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF international conference on computer vision. 13919--13929 jiang2021focal

    2021

  19. [19]

    and Rizzo, G

    Kanwal, N. and Rizzo, G. (2022). Attention-based clinical note summarization. In Proceedings of the 37th ACM/SIGAPP symposium on applied computing. 813--820 kanwal2022attention

    2022

  20. [20]

    Katebijahromi, N. (2021). Detection of Adverse Events in Pregnancy Using a Low-Cost 1D Doppler Ultrasound Signal. Ph.D. thesis, Emory University katebijahromi2021detection

    2021

  21. [21]

    Moyo fetal heart rate monitor

    [Dataset] Laerdal-Global-Health (2025). Moyo fetal heart rate monitor. https://shop.laerdalglobalhealth.com/product/moyo/. Accessed 23 July 2025 laerdal_moyo

    2025

  22. [22]

    Lin, Z., Liu, X., Wang, N., Li, R., Liu, Q., Ma, J., et al. (2024). Deep learning with information fusion and model interpretation for long-term prenatal fetal heart rate data. npj Women's Health 2, 31 lin2024deep

    2024

  23. [23]

    A., Levine, L

    McCoy, J. A., Levine, L. D., Wan, G., Chivers, C., Teel, J., and La Cava, W. G. (2025). Intrapartum electronic fetal heart rate monitoring to predict acidemia at birth with the use of deep learning. American Journal of Obstetrics and Gynecology 232, 116--e1 mccoy2025intrapartum

    2025

  24. [24]

    L., Chen, S., Ogata, Y., Suzuki, M., Kobayashi, Y., Toda, M., et al

    Mwakawanga, D. L., Chen, S., Ogata, Y., Suzuki, M., Kobayashi, Y., Toda, M., et al. (2024). Barriers and facilitators of fetal heart monitoring with a mobile cardiotocograph (ictg) device in underserved settings: An exploratory qualitative study from tanzania. PloS one 19, e0314812 mwakawanga2024barriers

    2024

  25. [25]

    Nageotte, M. P. (2015). Fetal heart rate monitoring. In Seminars in fetal and neonatal medicine (Elsevier), vol. 20, 144--148 nageotte2015fetal

    2015

  26. [26]

    P., Spilka, J., Stylios, C

    Oikonomou, V. P., Spilka, J., Stylios, C. D., and Lhotsk \'a , L. (2013). An adaptive method for the recovery of missing samples from fhr time series. In CBMS. 337--342 oikonomou2013adaptive

    2013

  27. [27]

    A., Redman, W

    Petrozziello, A., Jordanov, I., Papageorghiou, T. A., Redman, W. C., and Georgieva, A. (2018). Deep learning for continuous electronic fetal monitoring in labor. In 2018 40th annual international conference of the IEEE engineering in medicine and biology society (EMBC) (IEEE), 5866--5869 petrozziello2018deep

    2018

  28. [28]

    Rao, L., Lu, J., Wu, H.-R., Zhao, S., Lu, B.-C., and Li, H. (2024). Automatic classification of fetal heart rate based on a multi-scale lstm network. Frontiers in Physiology 15, 1398735 rao2024automatic

    2024

  29. [29]

    Ribeiro, M., Monteiro-Santos, J., Castro, L., Antunes, L., Costa-Santos, C., Teixeira, A., et al. (2021). Non-linear methods predominant in fetal heart rate analysis: a systematic review. Frontiers in Medicine 8, 661226 ribeiro2021non

    2021

  30. [30]

    Spilka, J., Chud \'a c ek, V., Kouck \`y , M., Lhotsk \'a , L., Huptych, M., Jank, P., et al. (2012). Using nonlinear features for fetal heart rate classification. Biomedical signal processing and control 7, 350--357 spilka2012using

    2012

  31. [31]

    abnormal

    Spilka, J., Georgoulas, G., Karvelis, P., Chud \'a c ek, V., Stylios, C. D., and Lhotsk \'a , L. (2014). Discriminating normal from “abnormal” pregnancy cases using an automated fhr evaluation method. In Artificial Intelligence: Methods and Applications: 8th Hellenic Conference on AI, SETN 2014, Ioannina, Greece, May 15-17, 2014. Proceedings 8 (Springer),...

    2014

  32. [32]

    Spyridou, K. K. and Hadjileontiadis, L. J. (2007). Analysis of fetal heart rate in healthy and pathological pregnancies using wavelet-based features. In 2007 29th annual international conference of the ieee engineering in medicine and biology society (IEEE), 1908--1911 spyridou2007analysis

    2007

  33. [33]

    Tahir Hussain Rizvi, S., Kanwal, N., Naeem, M., Cuzzocrea, A., and Coronato, A. (2025). Bridging simplicity and sophistication using glinear: A novel architecture for enhanced time series prediction. arXiv e-prints , arXiv--2501 tahir2025bridging

    2025

  34. [34]

    B., Kidanto, H., and Ersdal, H

    Urdal, J., Engan, K., Eftest l, T., Yarrot, L. B., Kidanto, H., and Ersdal, H. (2019). Noise and contraction detection using fetal heart rate and accelerometer signals during labour. In Proceedings of the 17th Scandinavian Conference on Health Informatics. vol. 12, 13 urdal2019noise

    2019

  35. [35]

    N., et al

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., et al. (2017). Attention is all you need. Advances in neural information processing systems 30 vaswani2017attention

    2017

  36. [36]

    N., Lee, A

    Wall, S. N., Lee, A. C., Niermeyer, S., English, M., Keenan, W. J., Carlo, W., et al. (2009). Neonatal resuscitation in low-resource settings: what, who, and how to overcome challenges to scale up? International Journal of Gynecology & Obstetrics 107, S47--S64 wall2009neonatal

    2009

  37. [37]

    Wang, J., Pei, S., Yang, Y., and Wang, H. (2024). Convolutional transformer-driven robust electrocardiogram signal denoising framework with adaptive parametric relu. Math Biosci Eng 21, 4286--4308 wang2024convolutional

    2024

  38. [38]

    Xu, L., Wang, G., Cao, Z., Chen, Q., Liu, G., and Wei, H. (2023). Research on multimodal deep learning based on cnn and vit for intrapartum fetal monitoring. In 2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) (IEEE), 4459--4464 xu2023research

    2023

  39. [39]

    Zhang, H., Liu, W., Shi, J., Chang, S., Wang, H., He, J., et al. (2022). Maefe: Masked autoencoders family of electrocardiogram for self-supervised pretraining and transfer learning. IEEE Transactions on Instrumentation and Measurement 72, 1--15 zhang2022maefe

    2022

  40. [40]

    Zhang, W., Yang, L., Geng, S., and Hong, S. (2023). Self-supervised time series representation learning via cross reconstruction transformer. IEEE Transactions on Neural Networks and Learning Systems zhang2023self

    2023

  41. [41]

    Zhao, Z., Deng, Y., Zhang, Y., Zhang, Y., Zhang, X., and Shao, L. (2019). Deepfhr: intelligent prediction of fetal acidemia using fetal heart rate signals based on convolutional neural network. BMC medical informatics and decision making 19, 1--15 zhao2019deepfhr

    2019

  42. [42]

    Zhao, Z., Zhang, Y., and Deng, Y. (2018). A comprehensive feature analysis of the fetal heart rate signal for the intelligent assessment of fetal state. Journal of Clinical Medicine 7. doi:10.3390/jcm7080223 jcm7080223

    work page doi:10.3390/jcm7080223 2018

  43. [43]

    Zhong, M., Yi, H., Lai, F., Liu, M., Zeng, R., Kang, X., et al. (2022). Ctgnet: automatic analysis of fetal heart rate from cardiotocograph using artificial intelligence. Maternal-Fetal Medicine 4, 103--112 zhong2022ctgnet

    2022