Abstract

Antepartum fetal heart rate (FHR) monitoring is crucial for assessing fetal well-being. Severe Fetal Distress, such as “Category III” FHR status, indicates potential fetal acidemia and demands timely clinical decisions. Early prediction of severe fetal distress before onset remains a critical unmet need, offering an opportunity for preventative interventions. This study aimed to develop a deep learning model capable of predicting impending severe fetal distress 5 minutes before onset. The dataset was derived from multicenter cohort (N=18,302 patients from 12 hospitals) consisting of a development (N=16,460) and an internal validation (N=1,842) cohort, along with an independent external validation cohort (N=510 patients from **, yielding 183,016 continuous monitoring windows). A deep learning model was developed to predict severe fetal distress 5 minutes before onset using FHR and uterine contraction (UC) signals. The severe fetal distress occurred in 2.68% (441/16,460) of the development cohort and 2.12% (39/1,842) of the internal validation cohort. In the external validation cohort, 46 patients (9.0%, 46/510) experienced severe fetal distress in 52 abnormal windows. The developed model predicted the severe fetal distress 5 minutes before onset, achieving an AUROC of 0.855 (95% CI, 0.799-0.902) in the internal validation cohort. The model also demonstrated excellent performance in the external validation cohort, simulating continuous real-world monitoring, with an AUROC of 0.917 (95% CI, 0.882–0.946). A deep learning model demonstrated high accuracy in predicting impending severe fetal distress in both internal and external cohorts. This tool shows promise as an early warning system to support clinical decision-making and facilitate preventative interventions to mitigate fetal compromise.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6181_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{SeoDon_RealTime_MICCAI2026,
        author = { Seo, Donghyeon AND Park, Hyungyu AND Kim, Ji Hoi AND Kim, Kwangsoo AND Kim, Sohee AND Park, Chan-Wook AND Park, Joong Shin AND Choi, Un Yung AND Yoon, Yu Gyeom AND Lee, Tae Kwan AND Lee, Hyung-Chul AND Lee, Garam AND Lee, Seung Mi},
        title = { { Real-Time Prediction of Impending Severe Fetal Distress Using Deep Learning on Fetal Heart Rate Monitoring } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents a deep learning approach to predict severe fetal distress from antepartum fetal monitoring signals. The main contribution is the implementation of a short-term prediction algorithm, where 5-minute segments of fetal heart rate and uterine contraction signals are used to predict the onset of severe fetal distress before it becomes clinically evident. The method is designed to work on continuous monitoring data and aims to provide an early warning signal based on physiological patterns.

    The authors also contribute a large multicenter study, including more than 18000 patients collected from multiple hospitals. This dataset is used for model development and internal validation, and an additional independent external cohort is used to evaluate the generalization of the method across different clinical settings. The study is formulated as a binary classification problem between normal cases and severe fetal distress, where distress is defined based on signal patterns annotated by clinical experts.

    In addition, the paper includes examples of the model outputs over time, showing how the predicted risk evolves before the onset of the event and how it behaves in normal cases.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The main strength of the paper is the use of a large multicenter dataset, including more than 18000 patients from multiple hospitals. This is an important aspect, as it increases the diversity of the data and supports the potential generalization of the model. Also is really important to have a dataset such as the one that this paper is proposing.
    • The model produces a continuous risk score over time, which allows the evolution of the prediction to be analyzed. This is relevant from a clinical perspective, because it can help to monitor changes in fetal condition instead of providing only a single binary decision.
    • The paper addresses a clinically relevant problem.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The study focuses on antepartum fetal monitoring, while much of the existing literature on CTG-based prediction is focused on intrapartum data. A more explicit discussion of this difference, including the specific challenges and implications of antepartum prediction, would help to better position the contribution.
    • The manuscript does not clearly describe the input data. It is not explicitly stated that the signals correspond to cardiotocography (CTG), including fetal heart rate and uterine contractions. This lack of clarity makes the methodology difficult to follow.
    • The development and internal validation data are described as coming from the “NIA (AI-Hub) fetal monitoring dataset”, but no reference or description of this dataset is provided. It is unclear whether this is a public dataset or a proprietary one.
    • Severe fetal distress is defined based on visual patterns in the same signals used as input, annotated by a limited number of clinicians. Given the known subjectivity and variability in CTG interpretation, this raises concerns about label reliability. This limitation is only briefly mentioned and deserves a more detailed discussion.
    • The model is trained using a single 5-minute window per patient. For abnormal cases, the window is selected immediately before the event, while for normal cases it is randomly selected from the recording. This creates an asymmetry between classes and does not reflect the progressive nature of fetal compromise. As a result, the model may learn to distinguish late-stage patterns from randomly sampled segments, rather than learning to predict the evolution toward distress.
    • The duration of the CTG recordings is not specified. This information is important to understand how much temporal context is available and to better interpret the 5-minute window strategy.
    • The manuscript reports an “optimal threshold” (0.294) in the figures, but the origin of this value is not clearly explained.
    • The use of a large multicenter dataset with external validation is a strength, but similar large-scale studies have been reported in the literature, especially in intrapartum CTG analysis. The use of both fetal heart rate and uterine contraction signals reflects standard clinical practice and is not a novel aspect. The proposed 5-minute prediction window is also not clearly justified.
    • The manuscript presents UMAP visualizations and Grad-CAM analysis, but these methods are not clearly introduced or explained. In particular, the UMAP projections do not show a clear separation between normal and abnormal cases, and the conclusions drawn from these visualizations are not sufficiently supported. This part of the analysis is difficult to interpret.
    • Some claims in the discussion are not supported by evidence. The manuscript suggests that a 5-minute prediction window would enable effective preventive interventions. However, this claim is not supported by references.
  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I recommend a weak reject for this paper. The work addresses an important clinical problem and is based on a large multicenter dataset with external validation, which are clear strengths. However, the main contributions are not fully convincing. While the dataset is strong, the other claimed contributions are less clear. The use of both fetal heart rate and uterine contraction signals reflects standard practice, and the proposed 5-minute prediction window is not sufficiently justified as a significant methodological or clinical advancement. In addition, there are several issues in the methodological description that make the paper difficult to follow. The input data is not clearly described, the dataset is not properly referenced, and the objective of the model is not clearly stated. Important details such as the duration of the recordings are missing. The windowing strategy is also a concern, as it does not reflect the progressive nature of fetal compromise and may bias the learning process. Furthermore, the labeling is based on subjective interpretation of the same signals used as input, which raises questions about reliability. The interpretability analysis and some claims in the discussion are not well supported, and overall the clarity of the manuscript could be improved. For these reasons, I believe the paper is below the acceptance threshold in its current form, although it addresses a relevant problem and could be improved with clearer methodology and stronger justification of its contributions.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    After reading the rebuttal, I believe that the authors addressed several of my main concerns, particularly regarding the intended scope of the work as a clinical-translation study, the rationale of the windowing strategy, and the interpretation of the evaluation setting. The clarification that the external validation was performed on continuous monitoring windows also helped to better understand the practical evaluation scenario. I still believe that the manuscript lacks some important methodological details and clarity. In particular, the description of the input data and dataset remains limited, and some aspects related to the temporal progression of fetal compromise are still not fully convincing from a physiological perspective. I also still think that the conclusions drawn from the UMAP and Grad-CAM analysis are not fully supported. However, considering the paper as a clinical-translation/application study rather than a methodological contribution, I believe that the large multicenter dataset, the independent external validation, and the reported predictive performance make the work relevant and potentially valuable for the MICCAI community.



Review #2

  • Please describe the contribution of the paper

    The paper presents a carefully done time series prediction study utilizing best in class existing methods using an impressing dataset, including a novel validation cohort developed by the authors.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.Impressing (and even partially innovative) dataset 2.Correct formulation of the prediction problem and avoidance of several very common statistical mistakes commonly done in these studies. 3.Use of best in class methods for all steps.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The metrics used for result presentation, while accuracy, sensitivity, and specificity are generally good, because rare event prediction is a problem with inherent class imbalance, I think NPV & PPV would have given the reader better sense of the actual clinical utility. 2.While methods are best in class, none is novel. 3.External validation dataset was not published and the model architecture is not elaborate for reproduction of the results.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not provide sufficient information for reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    While a very very neat work (well done, well explained), lack of innovation in methodology and insufficient data accessibility and model details for reproducibility are weakening it significantly. When I review for medical journals I’d strongly accept, but Those issues are IMO more important in venues like MICCAI on which methodology innovation and reproducibility are also expected and not only conducting a good clinical study

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper proposed a deep learning prediction method for fetal heart rate (FHR) monitoring. The author(s) also provided solid comparision study with multicenter data from real clinical scenarios. The result indicated a high accuracy for such clinical decision and could contribute to real-world intervention.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    From my perspective, the author(s) did really good job at: 1.conceive and complete a solid, detailed analysis on both the data selection & categorization and performance. 2.the author(s) demonstrated strong clinical feasibilty and thought carefully about its impact on real world application. 3.provides a scientific perspective on what factors impact the clinical feasibility and prediction accuracy.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    I do understand that there are some weaknesses which are probably come from the limitation of pages. I would really love to see the auther addressing the following issues: 1.The figures are not all self-explainable. For example, Fig 1 can be represented by word description instead of figure (since space is limited in MICCAI). Fig 2 is difficult to understand for me though I get the general idea. Would it be possible to add more description to Fig2? Especially in caption. Besides, although I understand the novelty of model architecture is not the point of this paper, since this is still a deep learning paper, could we extend subplot B in figure 2 so that it illustrates part 2.3 better.

    2.I am curious about a comparision between your data analysis and the potential bayesian analysis and statistics about part 3.Since you have the data and your result has a really low FP rate (Fig 3 B&D), it is essential to make comparison and say the deep learning is statistically useful.

    3.Fig 5, again, not really self-explainable and please enlarge middle figure (it is really small to see).

    4.Fig3 A & C, since we anyway have three combination of factors, it is better to demonstrate three AUROC curves on same diagram so we know the factor comparison.

    5.In general, I understand that there are limitation of pages, but this study lacks of parallel comparison. In other word, how would you 1D ResNet method compare to time-series analysis (or time-series DL method)? Would bigger network work better? Would transformer be better? I am pretty sure there are plenty of existing methods working on window-based prediction (not necessary in FHR but you get the idea). These are my suggestion if you want to extend this miccai paper.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Three major aspects of this study makes me consider this a valid work for MICCAI: 1.this paper did a thorough investigation on the clinical feasibility of using deep learning for FHR prediction. 2.The result is promising. 3.The author(s) provides explaination AI (e.g. causes & features for abnormal and normal condition).

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    From my point of view, this is still a neat and valuable work for FHR monitoring. The clinical value of this paper is huge. A more detailed description is desirable but can be extended to journal for better impact. I do wish to see more of these kind of works appearing in MICCAI, especially for CAI, instead of modifying networks, modules, etc.



Author Feedback

We thank the reviewers for their thoughtful feedback. We respond to the main issues below (R1/R2/R3 = Reviewers #1/#2/#3).

1.Scope and contribution [R1, R2, R3]: We thank the reviewers for their constructive comments, and would respectfully clarify our intended contribution as a clinical-translation study for antepartum continuous FHR/UC monitoring, predicting impending severe fetal distress 5 min before onset, as stated in Sec. 1.The independent external cohort (N=510; 183,016 continuous windows; Sec. 2.1) reflects a real-world setting, and FHR+UC fusion is empirically supported (Table 3: AUROC 0.855 vs. FHR-only 0.811, UC-only 0.786). We fully acknowledge that architecture is not the focus: as a 5-min FHR/UC window has low temporal dimensionality, we adopted the well-established 1D ResNet (Sec. 2.3, Fig. 2.B), robust in this regime. We agree with R3 that benchmarking other time-series methods is valuable future work.

2.Windowing strategy [R1]: Our task is intentionally framed as short-term impending-event prediction rather than full trajectory modeling. Sec. 2.2 specifies that “the 5-minute window immediately preceding the event was extracted” for abnormal cases, while for normal controls “one 5-minute window was randomly selected.” We gently note that external validation does not rely on this 1:1 sampling: “the remaining continuous record was segmented into 5-minute windows, excluding a washout period” (Sec. 2.2), yielding 183,016 continuous windows simulating real-world monitoring. The external AUROC of 0.917 (Sec. 3.2, Fig. 3.C) and sustained performance at 10/15/20-min lead times (Table 4: 0.864/0.851/0.870) may help mitigate concerns of late-stage shortcut learning.

3.Outcome definition and label reliability [R1]: We agree that inter-/intra-observer variability is a genuine limitation, acknowledged in Sec. 4.3: “the labeling…was performed by a single resident, which may introduce subjective bias, although the criteria were strictly defined.” Our target is clinically meaningful precursor FHR patterns, not direct prediction of acidemia: severe fetal distress was operationalized as “prolonged deceleration, repetitive late deceleration, repetitive variable deceleration, bradycardia, or absent variability” (Sec. 2.1, Table 1), mirroring Category III criteria.

4.Evaluation metrics and threshold [R1, R2]: We agree that PPV/NPV are important for rare-event prediction. The manuscript reports AUROC as the primary discrimination metric and threshold-based sensitivity/specificity; Fig.3 confusion matrices allow PPV/NPV to be derived. Regarding threshold question, Sec.2.4 states that operating threshold was selected by the Youden Index on internal validation and then applied directly to external validation, without external tuning.

5.Clinical utility and scope of claims [R1]: We agree that prospective validation is needed to prove clinical effectiveness. Our claim is therefore discriminative early-warning capability, not proven outcome improvement; Sec.4.3 states that real-time workflow/outcome impact remains unmeasured.

6.Reproducibility and data availability [R1, R2, R3]: The NIA(AI-Hub) development data and hospital external data are clinical datasets subject to national data-access policies and IRB/hospital restrictions and were not presented as openly released. Fig. 2 and Sec. 2.3 describe the architecture and training in detail. We are encouraged that R3 found the description “clear and detailed.”




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The manuscript was reviewed by three expert reviewers. While acknowledging merits in the proposed work, the reviewers also raised some concerns about methodological novelty and validation. Such concerns should be addressed to strengthen the value of the work.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The reviewers’ concerns have been successfully addressed in the rebuttal, making the work a solid contribution especially for what concerns clinical translation.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The majority of reviewers recommend acceptation of the paper. Concerns are largely addressed in the rebuttal. Despite concerns related to methodological explanation and clarity, clinical Translation/application nature of the paper are convincing and there is strong multicenter dataset with external validation. The study is well conducted and has high clinical relevance.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The reviewers raised that the rebuttal addressed several concerns. However, I would recommend to address those issues in the camera-ready version.



back to top