CTG-FRAME: development and validation of a multi‑modal deep learning for intrapartum escalation support and severe fetal acidemia prediction

Abstract

Background

Cardiotocography is important for continuous intrapartum fetal surveillance, yet its clinical effectiveness is limited by high inter-observer variability and a false-positive burden of approximately 60%. These limitations contribute to unnecessary intervention while delaying recognition of fetal acidemia. For clinical use, decision-support systems must not only achieve high accuracy but also define their role in guiding escalation of care.

Methods

We developed and evaluated CTG-FRAME (CardioTocoGraphy–Fetal Risk Assessment and Monitoring Engine), an end-to-end deep learning decision-support framework for identifying risk of severe fetal acidemia (umbilical arterial pH < 7.05) using fetal heart rate and uterine contraction signals. The framework integrates self-supervised momentum-contrast pre-training, a hybrid ResNet–temporal convolutional encoder, bidirectional cross-attention, and a Transformer to capture global temporal context. The model was trained and internally validated on the CTU-UHB dataset (n = 552; ~8% acidemia) using three-fold cross-validation. External transferability and calibration were assessed in a separate multicentre cohort (SPaM; n = 300; ~20% acidemia) not used during model development. CTG-FRAME outputs a probability of acidemia, estimates continuous pH severity, and identifies signal segments contributing to each prediction. Training employed a difficulty-aware multi-stage curriculum to address severe class imbalance. Model performance was evaluated using AUROC, sensitivity, specificity, and agreement measures.

Results

On CTU-UHB, the model achieved an AUROC of 0.888, with sensitivity 0.850 and specificity 0.964 at the predefined operating threshold (≈0.52). External evaluation of SPaM showed predicted abnormality rates consistent with cohort prevalence without re-tuning, indicating calibration preservation across sites. Ablation analysis demonstrated that the staged curriculum was necessary to prevent collapse to majority-class prediction, and that the self-supervised pre-training substantially improved pH estimation accuracy.

Conclusions

CTG-FRAME is designed to facilitate the escalation of clinical review rather than autonomous clinical action by defining a clear decision-analytic operating region and generating interpretable, context-linked evidence. This work advances obstetric artificial intelligence by making clinical intent, operating trade-offs, and workflow roles explicit, moving beyond accuracy toward verifiable bedside utility.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03651-6.

Background

Cardiotocography (CTG) is a bedside monitoring technique employed during pregnancy labor to assess fetal well-being by continuously recording the fetal heart rate (FHR) and uterine contractions (UC). Clinicians evaluate changes in FHR relative to contractions to determine fetal tolerance of labor. CTG serves as the standard modality for continuous intrapartum surveillance from approximately 28 weeks of gestation until delivery [1]. However, its clinical effectiveness is limited by substantial inter-observer variability and a false-positive alarm rate estimated at approximately 60%, which can lead to unnecessary interventions and delayed responses to genuine fetal compromise [2, 3].

These trade-offs are inherently harm-asymmetric. Failure to detect severe acidemia results in greater clinical consequences than performing an unnecessary operative delivery. The high rate of false-positive alarms can erode clinical vigilance and undermine clinicians’ confidence in the monitoring process. Therefore, an effective decision-support system must not only achieve high overall accuracy but also clearly specify the clinical decision it is intended to support, characterize harm-asymmetric operating-point trade-offs, and demonstrate how its outputs integrate into intrapartum workflows without replacing clinical judgment.

CTG signals (FHR and UC) are inherently noisy, subject to sensor dropout, motion artifacts, and baseline drift, rendering signal pre-processing a prerequisite for reliable interpretation (Fig. 1). Even under optimal conditions, expert visual assessment yields considerable inter-observer disagreement, with concordance rates as low as approximately 29%, sustaining a persistent clinical dilemma between over-intervention and missed acidemia, both of which carry well-documented maternal and neonatal risks [5, 6].

Fig. 1.

FHR and UC signals from a sample CTG recording of the CTU-UHB dataset [4]. Representative fetal heart rate (FHR) and uterine contraction (UC) signals from a single cardiotocography (CTG) recording in the CTU-UHB dataset. Blue traces show the raw signals as typically encountered in routine intrapartum monitoring, including noise and artifacts. Orange traces show the same signals after cleaning using the CTG-FRAME preprocessing pipeline, and green traces show the normalized signals used as model inputs.(a) FHR signal illustrating physiological variability and common artifacts in raw recordings, such as transient signal dropouts. (b) Corresponding UC signal from the same recording, demonstrating contraction timing and intensity, which provides essential context for interpreting FHR changes. For both panels, the x-axis represents time (samples), and the y-axis represents fetal heart rate (beats per minute) or uterine contraction intensity (mmHg or relative units), respectively

Automated CTG analysis has yielded promising but fragmented progress (6). Early approaches relied on hand-crafted features and conventional machine learning, which proved insufficient for capturing the subtle, non-linear physiology embedded in raw signals [7–9]. Contemporary deep learning methods address this limitation by enabling end-to-end learning that jointly encodes local morphological features and global temporal context [10, 11]. Nevertheless, a central clinical gap persists: decisive intrapartum judgments frequently depend on inter-signal timing. Specifically, the latency of FHR decelerations relative to contraction onset, which simple fusion approaches handle inadequately [12, 12]. Clinicians interpret FHR changes in relation to UC rather than as independent signals, and accuracy alone remains insufficient unless results are anchored to a decision-analytic operating point aligned with clinical escalation workflows.

Progress is further constrained by data quality and availability. Pre-processing standards vary widely across studies, impeding direct comparison. The field remains anchored to small public datasets (e.g., CTU-UHB, n = 552), while larger real-world collections remain inaccessible [13, 14]. Widely used benchmarks such as the UCI CTG dataset provide engineered features rather than raw signal data, precluding evaluation of approaches that preserve subtle morphology and cross-modal relationships. Severe class imbalance is an intrinsic characteristic of available datasets, given the rarity of severe acidemia; naive single-stage training consequently produces models biased toward the majority class. Common remedies, including SMOTE oversampling and generative adversarial network (GAN)-based synthesis, are frequently applied without clinical rationale and risk introducing physiologically implausible signals [15, 16]. Evaluation practice further limits clinical translation: many studies rely exclusively on internal cross-validation of a single benchmark [5, 17]. Taken together, these constraints underscore the need for multi-task outputs: encompassing risk classification, severity estimation, and interpretable evidence, reported at clinically relevant operating thresholds rather than threshold-free discrimination metrics alone.

Within this landscape, recent approaches demonstrate genuine promise yet remain fragmented. Several studies report strong headline metrics on CTU-UHB and explore directions such as lightweight deployment, multi-modal fusion, and the inclusion of clinical covariates. However, the majority address a single classification task, rely solely on internal validation, and do not specify an intended operating threshold at which a clinician could act (27–30). Three practical requirements for clinical translation therefore remain only partially met: (i) explicit modeling of FHR–UC inter-signal timing; (ii) multi-task outputs encompassing risk classification, severity estimation, and interpretable evidence aligned to bedside verification; and (iii) performance reporting at a clinically intended threshold, rather than threshold-free discrimination metrics alone [18, 19].

This paper presents CTG-FRAME (CardioTocoGraphy–Fetal Risk Assessment and Monitoring Engine), an end-to-end deep learning framework designed to address these interdependent challenges within a unified pipeline. Rather than optimizing components in isolation, CTG-FRAME co-designs signal pre-processing, multi-modal representation learning, and training strategy so that each element is optimized in concert, reflecting the clinical principle that utility emerges from component synergy rather than any single module [5, 15, 20]. From a clinical decision-support perspective, CTG-FRAME integrates five elements that prior work has explored separately but which are necessary in combination: (i) self-supervised momentum contrast (MoCo) pre-training to learn robust physiological representations from unlabeled CTG windows and mitigate label scarcity [21]; (ii) a hybrid multi-modal encoder pairing a ResNet–TCN backbone for local morphology with bidirectional FHR⇄UC cross-attention for clinically critical inter-signal timing; (iii) a difficulty-aware, multi-stage curriculum that stabilizes optimization under severe class imbalance and prevents majority-class collapse [22, 23]; (iv) a clinically driven multi-task objective producing binary acidemia classification, continuous pH regression with uncertainty quantification, and event segmentation for clinicians’ actionability; and (v) a dual-validation strategy coupling internal cross-validation on CTU-UHB with external calibration on SPaM as an initial transferability assessment [4, 24].

Each design choice is clinically informed and empirically justified through ablation analysis. The difficulty-aware curriculum was essential for achieving a balanced, safety-aligned operating region; naive single-stage training produced majority-class collapse. Removing self-supervised pre-training nearly doubled pH estimation error (RMSE: 0.563 to 1.067; +89%), establishing that the severity estimation head depends critically on robust foundational representations. At the intended clinical threshold—calibrated to reflect obstetric harm asymmetry—the pooled model achieved a sensitivity of 0.850 and specificity of 0.964 (AUROC 0.888; balanced accuracy 0.907; Youden’s J 0.814), consistent with intrapartum escalation practice. On external calibration assessment, classification rates approximated cohort prevalence without re-tuning, indicating reasonable calibration and supporting early transferability. This performance profile reflects a deliberately conservative, escalation-support operating point.

The remainder of this paper details the full pipeline, presents internal and external calibration assessment results at the intended operating region alongside interpretability analyses, and discusses translational implications and future research directions.

Methods

Reporting guideline statement

Reporting follows the TRIPOD+AI guideline, as CTG-FRAME is developed as an AI-based clinical prediction model for severe fetal acidemia risk. Relevant STARD-AI items were additionally considered, as the binary classification output can also be interpreted as an AI-centered index test evaluated against umbilical arterial cord pH as the reference standard. A completed TRIPOD+AI/STARD-AI compliance checklists are provided in Supplementary Table S1 to S3.

Intended use

CTG-FRAME is intended for use by midwives and obstetricians in intrapartum CTG surveillance settings during singleton term or near-term labors consistent with the characteristics of the development datasets. The system is designed to support the escalation of clinical review, including to senior clinician review where locally appropriate, by identifying traces at elevated risk of severe fetal acidemia and providing temporally localized signals for clinician verification. CTG-FRAME is not intended for autonomous clinical decision-making, nor for use in antepartum CTG interpretation, preterm gestations, prolonged second-stage labor, or prediction of neonatal prognosis.

System overview

CTG-FRAME is a modular decision-support system designed to analyze intrapartum CTG and generate temporally localized risk estimates for severe fetal acidemia. The architecture supports reproducible CTG signal preprocessing, segmentation, model inference, and attribution. A high-level overview of the system workflow is provided in Fig. 2, with detailed implementation and engineering specifications described in Supplementary Methods S1.

Fig. 2.

CTG-FRAME workflow

Datasets and clinical endpoints

A two-stage evaluation design was employed to ensure that findings from model development were evaluated against genuinely unseen data. All training and internal performance assessment were conducted on the CTU-UHB intrapartum benchmark; a separate, held-out corpus (SPaM) was reserved exclusively to assess the transferability of model outputs across sites (Table 1). Critically, no external data influenced model development or the selection of the decision threshold [4, 24]. This separation mitigates the overfitting risk inherent to approaches that rely solely on internal validation and provides a reproducible basis for evaluating the generalizability of CTG-FRAME.

Table 1.

Dataset for development, internal validation, and external calibration assessment

Criterion	CTU-UHB	SPaM
Records available	552	300
Severe acidemia, n (%)	44 (≈8%)	60 (20%; by design)
pH threshold	≤7.05 (consistent throughout)	≤7.05 (consistent throughout)
Comparator group	All pH > 7.05 cases	pH > 7.25 (enriched non-acidemic)
Gestational age	Not individually reported; intrapartum cohort	≥36 weeks (singleton)
Signal quality inclusion	≥40 min usable within final 90 min	>60 min; <15% signal loss; delivery within 10 min
Second-stage exclusion	>30 min excluded	Not specified in public documentation
Reference standard	Arterial cord pH (validated)	Arterial cord pH (validated)
Role in study	Development and internal validation	External post-training calibration check only

Primary outcome and reference standard

The primary outcome is the prediction of severe fetal acidemia, defined by umbilical cord arterial pH < 7.05. Umbilical arterial blood gas analysis at birth was used as the reference standard, as it is a valid, objective, and reproducible reference standard, routinely collected, and temporally proximal to intrapartum events [25]. Cord blood pH reflects systemic fetal acid–base status at the time of delivery, providing an established quantitative marker of acute intrapartum compromise. However, it is important to distinguish this from tissue-level hypoxia or metabolic acidosis within specific organs, which cannot be directly inferred from cord blood measurements alone [3]. Accordingly, the outcome captures acute acidemia risk at birth, rather than the broader spectrum of fetal oxygenation dynamics or downstream clinical injury.

Signal context and data integrity

CTG records two continuous signals: FHR and UC. Clinicians interpret these in relation to one another, assessing, for example, whether a fall in heart rate (a deceleration) occurs late relative to contraction onset, a pattern associated with greater clinical concern. Because these signals are acquired continuously and are susceptible to noise, dropouts, artifacts, and baseline drift, the integrity of data preparation and partitioning is consequential for both model accuracy and clinical safety.

CTU-UHB (development and internal validation dataset)

CTU-UHB comprises 552 intrapartum CTG recordings acquired at the obstetrics ward of the University Hospital Brno (Czech Republic) between 2009 and 2012, sampled at 4 Hz [26]. Included recordings were selected from those ending as close as possible to delivery and containing at least 40 minutes of usable signal within the final 90 minutes of labor. Recordings with a second stage of labor exceeding 30 minutes were excluded due to recognized interpretive difficulty during this phase. FHR was recorded via external Doppler ultrasound (n = 412), scalp electrode (n = 102), or both (n = 35); the acquisition method was unavailable for 3 records. UC was recorded using commercially deployed CTG systems (STAN S31, Neoventa Medical; Avalon FM40 and FM50, Philips Healthcare) as part of standard care. All records are linked to umbilical arterial blood gas analysis at delivery; 44 cases met the pH ≤ 7.05 threshold for severe fetal acidemia (~8%), reflecting the real-world rarity of this outcome.

Available clinical variables not incorporated into model development include maternal characteristics (age, parity, gravidity), delivery characteristics (mode and duration of delivery, meconium-stained fluid, acquisition method), fetal characteristics (sex, birth weight), and neonatal indicators (Apgar scores, extended cord blood gases, and early neonatal care measures). Given the pH reference standard, these variables were excluded to avoid introducing heterogeneity related to postnatal care, documentation variability, and downstream decision-making. In addition, several of these variables reflect post hoc clinical assessment and management rather than physiological signal behaviour during labor. Consistent with this design rationale, ablation analysis (Table 2) demonstrated that augmenting the model with externally derived or hand-crafted features did not improve and, in some configurations, degraded performance, suggesting that, within this dataset, poorly harmonized auxiliary inputs may introduce noise rather than a complementary signal.

Table 2.

Aggregated cross-validation performance of the CTG-FRAME

Task	Metric	Fold 1	Fold 2	Fold 3	Mean of Folds (± SD)	Pooled Performance [95% Confidence Intervals]
Binary Fetal Acidemia	AUROC	0.967	0.992	0.962	0.974 (±0.013)	0.888 [0.832–0.937]
	Sensitivity (Recall)	0.926	0.968	0.897	0.930 (±0.029)	0.850 [0.771–0.909]
	Specificity	0.927	0.969	0.9	0.932 (±0.028)	0.964 [0.950–0.986]
	F1-Score	0.926	0.969	0.897	0.931 (±0.030)	0.853 [0.800–0.909]
	Youden’s J Statistic	0.853	0.937	0.797	0.862 (±0.058)	0.813 [0.736–0.881]
	Composite Score	0.595	0.628	0.5753	0.599 (±0.022)	0.536 [0.483–0.578]
	Accuracy	0.927	0.969	0.898	0.931 (±0.029)	0.940 [0.922–0.964]
	Balanced Accuracy	0.926	0.968	0.898	0.931 (±0.029)	0.907 [0.868–0.940]
	Threshold	0.540	0.517	0.6104	0.556 (±0.040)	0.517 [0.517–0.582]
Continuous pH Regression	RMSE (pH units)	0.502	0.323	0.602	0.476 (±0.115)	0.563 [0.516–0.612]
	Concordance Corr. Coeff. (CCC) pH	0.905	0.959	0.849	0.904 (±0.045)	0.823 [0.780–0.856]
CTG Event Segmentation	Dice Score (Macro Avg.)	0.382	0.409	0.583	0.458 (±0.089)	0.333 [0.269–0.395]
	mIoU (Macro Avg.)	0.318	0.318	0.549	0.395 (±0.109)	0.277 [0.238–0.327]

SPaM (external post-training calibration dataset)

The SPaM dataset comprises 300 intrapartum CTG recordings collected across three centres: Lyon (France), Brno (Czech Republic), and Oxford (UK) [24]. Each site contributed 100 cases, including 20 cases with umbilical arterial cord pH < 7.05 and 80 cases with pH > 7.25, resulting in an overall severe acidemia prevalence of approximately 20%. Recordings were acquired as part of routine clinical practice, sampled at 4 Hz, restricted to singleton pregnancies ≥ 36 weeks with validated arterial cord blood gas measurements, >60 minutes duration, <15% signal loss, and ending within 10 minutes of delivery. The deliberate prevalence enrichment provides a rigorous test of calibration across heterogeneous clinical sites. The SPaM dataset was used exclusively for post-training calibration assessment. The decision threshold derived from CTU-UHB was applied unchanged, such that SPaM evaluation reflects preliminary evidence for cross-site probability calibration and transferability, rather than re-optimized discrimination.

Signal pre-processing

Raw CTG recordings cannot be analyzed directly due to noise, signal dropout, and device-dependent variability. Therefore, a clinically informed, fully reproducible preprocessing workflow was developed to convert paired FHR and UC signals into standardized inputs suitable for model training and inference (Fig. 3). This workflow was designed to mirror how clinicians visually review CTG traces in practice, discounting artifact, tolerating brief monitoring interruptions, and interpreting FHR responses in relation to contractions over time. A high-level summary is provided here; full mathematical definitions and implementation details are reported in Supplementary Methods S1.

Fig. 3.

Data preparation and input construction for CTG-FRAME (advanced pathway). Overview of the advanced data preparation workflow, designed to reflect routine clinical CTG interpretation. Raw FHR and UC signals are first reviewed for physiologically implausible values and transient artifacts, analogous to excluding obvious signal loss or probe displacement at the bedside. Short gaps consistent with brief monitoring interruptions are cautiously interpolated, while longer signal absences are preserved as missing data. The cleaned traces are then examined across overlapping time windows to reflect continuous bedside review, with windows stratified by early and late labor phases to account for changing physiological context as delivery approaches. These windows are assembled into model-ready inputs that retain the temporal relationship between FHR and UC, reflecting how clinicians assess heart rate responses relative to uterine contractions. Hand-crafted features and metadata are generated solely for ablation analyses and visualisation and are not used by the final model

Two approaches were considered. The simpler approach directly fills in missing or corrupted parts of the signal, treating these gaps as small interruptions in otherwise reliable data. This method is straightforward but can unintentionally blend true physiological patterns with noise. The advanced approach used in the final model takes a different sequence of steps. Instead of immediately filling in missing or corrupted parts of the signal, it first identifies and removes segments that are clearly unreliable (for example, due to movement or poor sensor contact). Only after these unreliable segments are excluded are short gaps filled in. This difference in order is important: the simpler approach fills gaps directly, which can mix noise with true physiological changes. In contrast, the advanced approach cleans the signal before filling gaps, preventing artifacts from being blended into the reconstructed signal. In practice, this preserves clinically meaningful patterns and yields more reliable results (Table 1).

Within the advanced approach, physiologically implausible samples and transient artifacts are identified and masked prior to any reconstruction. For FHR, samples UC signals, artifact detection is performed relative to the statistical distribution of each individual trace, identifying samples that deviate markedly from the typical contraction pattern observed within that recording.

Importantly, UC preprocessing is performed in a trace-relative, device-agnostic manner and does not assume absolute pressure units. Because CTG datasets do not uniformly specify whether uterine activity was measured with an external tocodynamometer or an internal intrauterine pressure catheter, UC amplitude is treated as a relative signal describing temporal contraction patterns rather than as true intrauterine pressure in mmHg. All preprocessing operations are therefore designed to be invariant to monitor calibration and zeroing.

Short signal dropouts are reconstructed only for gaps shorter than 60 seconds to improve robustness against transient artifacts such as brief transducer displacement or maternal movement. Gaps of 60 seconds or longer remain masked and are never reconstructed. The 60-second threshold was selected deliberately as a conservative upper bound for reconstruction: long enough to accommodate brief, non-physiological signal interruptions, yet short enough to avoid synthesizing extended segments that could plausibly contain sustained or evolving pathological patterns. Sensitivity analysis across shorter reconstruction thresholds (15, 30, and 45 seconds) demonstrated that the model remained capable of discriminating severe fetal acidemia, but with increased variability and reduced agreement in continuous pH estimation, consistent with fragmentation of physiologically meaningful temporal patterns (Supplementary Table S4). These findings support the use of a 60-second threshold to preserve temporal coherence and stability of representation, rather than maximize reconstruction coverage.

Importantly, signal reconstruction in CTG-FRAME is intended to support model robustness to incomplete data, not to restore or reinterpret physiological events. From a clinical perspective, any model alert arising during or immediately adjacent to reconstructed signal segments should prompt review of the raw CTG trace and signal quality, rather than being interpreted as confirmation of underlying physiology.

All reconstructed samples are explicitly flagged via a binary missingness indicator that is retained throughout model training and inference, allowing the model to condition its predictions on data availability rather than treating reconstructed values as equivalent to observed measurements. This design ensures that potential uncertainty introduced by reconstruction is preserved rather than concealed. As a result, missingness itself is modeled as informative, and reconstructed segments do not override physically observed signal evidence.

Full details for reproducibility are available in Supplementary Methods S1.

Model architecture

CTG-FRAME is a hybrid, end-to-end sequence model designed to reflect the hierarchical way clinicians interpret CTG. In routine practice, clinicians assess short-term morphological features, such as the shape, depth, and timing of decelerations, while also considering longer-term temporal context, including whether abnormalities are accumulating, persisting, or resolving as labor progresses. CTG-FRAME mirrors this process by combining local pattern analysis with global temporal integration across the full CTG recording.

The architecture comprises three coordinated components (Fig. 4). First, self-supervised pre-training, a learning approach that uses unlabeled CTG recordings to learn general signal structure before outcome labels are introduced, is used to acquire robust signal representations, reducing dependence on scarce annotated data. Second, a hybrid convolutional backbone combines residual connections (which stabilize deep learning) with temporal convolution (which captures patterns over different time scales) to extract clinically meaningful features from FHR and UC signals. These features are integrated using bidirectional cross-attention, a mechanism that allows the model to explicitly relate changes in FHR to the timing of UC, analogous to how clinicians assess decelerations relative to contractions. Third, a Transformer encoder aggregates evidence across overlapping time windows to produce a case-level risk assessment, enabling the model to account for temporal evolution across the recording rather than isolated events.

Fig. 4.

Model architecture. Each FHR and UC signal is first processed independently to extract local morphological patterns. Information from the two signals is then integrated using bidirectional cross-attention, enabling interpretation of heart rate changes in relation to contraction timing. A Transformer encoder aggregates evidence across all time windows to form a case-level representation, which supports risk classification, continuous pH estimation with uncertainty, and localization of clinically relevant signal segments

Together, these components allow CTG-FRAME to jointly model short-term signal morphology and longer-range temporal context in a manner aligned with clinical reasoning. Full architectural specifications, layer configurations, and training details are provided in Supplementary Methods S2, and definitions of technical terms are summarized in Supplementary Appendix S1 (Glossary).

Self-supervised pre-training

Severe fetal acidemia is a relatively rare outcome, which means that labeled CTG data that reflect sufficient normal and severe samples are inherently limited. To improve learning under this constraint, the CTG-FRAME encoder, the component responsible for extracting meaningful patterns from raw FHR and UC signals, is first trained before any outcome labels are used.

This initial training uses self-supervised contrastive learning, a learning strategy in which the model is exposed to large volumes of unlabeled CTG recordings and learns the general structure of normal and abnormal signal patterns without being told the clinical outcome. In practical terms, this allows the encoder to learn what typical CTG variability, decelerations, accelerations, and noise look like before it is asked to distinguish high-risk from low-risk cases. As a result, subsequent outcome-labeled training is more stable and less prone to overfitting to the small number of severe cases.

Specifically, to address label scarcity and expose the encoder to broad physiological variability, the model is pre-trained on unlabeled 240-second CTG windows using Momentum Contrast v2 (MoCo v2). MoCo v2 is a self-supervised learning framework that teaches the encoder to recognize when two slightly altered views of the same CTG segment represent the same underlying physiological pattern, while distinguishing them from different segments [27, 28]. This approach has been shown to produce robust, generalizable signal representations in settings where annotated data are scarce, including medical time-series and imaging tasks [29]. Full implementation details are provided in Supplementary Methods S2, and key technical terms are defined in Supplementary Appendix S1 (Glossary).

Hybrid ResNet–TCN backbone, bidirectional cross-attention, and transformer encoder

The architecture is organized into three sequential stages that reflect how clinicians interpret CTG: local pattern recognition, integration of FHR and UC, and global assessment of temporal evolution across labor.

• Local feature extraction (ResNet–TCN): Each 240-second segment of FHR and UC data is first processed to identify clinically meaningful short-term patterns, such as baseline characteristics, variability, and deceleration morphology. This is achieved using a convolutional encoder (ResNet-TCN) that combines residual connections with temporal filtering, allowing the model to capture signal features at multiple time scales while remaining stable during training. Separate encoders are applied to FHR and UC signals, reflecting their distinct physiological roles. These encoders are initialized using the self-supervised pre-training described above, enabling them to start from representations that already capture general CTG signal structure rather than relying solely on scarce labeled outcomes.

• Cross-modal fusion (FHR⇄UC cross-attention): In clinical practice, FHR patterns are interpreted explicitly in relation to UC, for example, determining whether a deceleration occurs early, coincident with, or late relative to a contraction. To model this relationship, CTG-FRAME integrates information from the FHR and UC streams using bidirectional cross-attention, a mechanism that allows each signal to condition its interpretation on the other. This structured fusion enables the model to represent clinically decisive timing relationships rather than treating FHR and UC as independent or merely co-occurring signals. In ablation analyses, this approach outperformed simple signal concatenation, supporting the need for an explicit, learnable interaction between modalities.

• Global contextualization (Transformer encoder): Following local feature extraction and cross-signal integration, information across all time windows is aggregated to support a case-level assessment. This is performed using a Transformer encoder, which integrates evidence across the full CTG recording and captures longer-range temporal patterns, such as progressive baseline change or accumulating deceleration burden. This step ensures that predictions are informed by temporal trajectories across labor, rather than isolated events within individual windows. The resulting global representations support the model’s multiple outputs, including binary risk classification, continuous cord-pH estimation with uncertainty, and localization of clinically relevant signal segments.

Together, these architectural components address complementary limitations of prior CTG analysis approaches. Self-supervised pre-training improves representation quality under label scarcity; bidirectional cross-attention enables clinically meaningful interpretation of FHR relative to contractions; and global temporal aggregation aligns predictions with the longitudinal nature of intrapartum assessment. Full architectural specifications, mathematical formulations, and training configurations are provided in Supplementary Methods S2, with technical terms defined in Supplementary Appendix S1 (Glossary).

Training strategy

Training CTG-FRAME integrates two interdependent components designed to reflect clinical priorities under real-world constraints: (i) a multi-task learning objective that mirrors the multi-dimensional decisions made during intrapartum assessment, and (ii) a difficulty-aware, multi-stage training curriculum that stabilizes learning under the severe class imbalance characteristic of CTG data (Fig. 5). Together, these components support a balanced, safety-aligned operating point suitable for clinical decision support.

Fig. 5.

Difficulty-aware multi-stage training curriculum. Overview of the CTG-FRAME training strategy. Training progresses from self-supervised pre-training to a three-stage supervised curriculum (W0–W2). The curriculum uses adaptive sampling and task weighting to focus learning on clinically relevant and challenging cases, addressing severe class imbalance and supporting a deployment-relevant operating point

Multi-task learning objective

Clinical intrapartum assessment is inherently multi-dimensional: clinicians evaluate whether fetal compromise is present, assess its severity, and identify the signal patterns driving concern. CTG-FRAME reflects this structure by producing, in a single forward pass, three complementary outputs: (i) a binary probability of severe fetal acidemia, (ii) a continuous estimate of cord-blood pH with associated uncertainty, and (iii) per-window event predictions to support interpretability and trace review.

These tasks are trained together using a stage-aware weighting strategy within the learning objective, which determines how strongly each task influences model training at different stages. This allows learning priorities to shift over time, preventing common low-risk patterns from dominating early training and ensuring that rare but clinically critical outcomes are adequately represented.

• Classification (acidemia probability): The binary classification task is trained using a class-weighted focal loss, which explicitly reflects the asymmetric clinical risk of missed severe acidemia compared with false alarms. In practice, this encourages the model to focus learning on difficult and minority-class cases rather than being dominated by the large number of normal CTG recordings. The degree of emphasis on harder cases is adjusted across training stages as part of the curriculum strategy described below. Full mathematical definitions are provided in Supplementary Methods S2.

• pH Regression (severity estimation with uncertainty): To estimate the severity risk of fetal acidemia, the model predicts both a mean cord-pH value and an associated uncertainty. Training uses a probabilistic regression objective that penalizes over-confident errors more strongly than uncertain ones, encouraging calibrated severity estimates rather than point predictions alone. This formulation supports a more cautious approach to borderline cases and aligns with the inherent uncertainty in intrapartum assessment. Technical details are provided in Supplementary Methods S2.

• Event Segmentation (interpretability support): To support clinical interpretability, the model also produces per-window predictions of clinically relevant signal events, such as late and prolonged decelerations. These predictions are trained using a loss function designed to be robust to rare foreground events, reflecting the low prevalence of pathological patterns within long CTG recordings. This task does not drive clinical decisions directly but provides contextual information to help clinicians understand why a trace was flagged.

Difficulty-aware, multi-stage curriculum

Typical naïve single-stage training, often employed in training an AI model, often leads to a default toward predicting normality, since most CTG datasets are dominated by normal cases [30, 31]. To address this, CTG-FRAME employs a three-stage curriculum (W0–W2) in which sampling strategies and task weights are progressively adjusted. Early stages prioritise stable learning and broad pattern recognition, while later stages increasingly emphasise difficult, clinically high-risk cases and severity estimation. This staged approach was necessary to prevent majority-class collapse and achieve a clinically meaningful balance between sensitivity and specificity. Full curriculum schedules and weighting schemes are described in Supplementary Methods S2.

W0—Warm-up to establish astable signal representation. In the initial stage, the model is exposed to a balanced mixture of normal and abnormal CTG patterns while the pre-trained signal encoders are kept fixed. This protects previously learned physiological representations and allows the newly introduced prediction components to stabilize their interpretation of CTG features. At this stage, training emphasizes robustness rather than fine discrimination, preventing early overfitting and ensuring a stable foundation for subsequent learning.

W1—Refinement to consolidate normal–abnormal boundaries. In the second stage, training emphasis shifts toward more challenging cases while maintaining encoder stability. The model is encouraged to sharpen its distinction between normal physiological variation and early signs of compromise, improving its ability to avoid false alarms without sacrificing sensitivity. This stage strengthens the model’s internal representation of normal labor physiology, which is essential for clinically meaningful specificity.

W2—Full fine-tuning to focus on the hardest clinical cases: In the final stage, all components of the model are allowed to adapt jointly, and training concentrates on the most difficult and informative cases (those most likely to be misclassified). This stage integrates earlier learning into a coherent end-to-end solution, refining both sensitivity to severe fetal acidemia and overall balance between missed detections and false alerts. By the end of this stage, the model converges on a stable operating point appropriate for deployment-oriented decision support.

Across training stages, the proportion of abnormal cases presented during training (formalized as the Target Positive Fraction - TPF) is adaptively adjusted based on the observed error profile in the preceding stage, specifically the relative balance of false-negative and false-positive errors. When false negatives are over-represented, indicating missed severe acidemia, subsequent training stages increase exposure to abnormal and borderline cases; conversely, when false positives dominate, sampling is rebalanced toward normal physiology. This feedback mechanism steers learning toward safer clinical trade-offs, progressively aligning sensitivity and specificity with the intended intrapartum escalation role rather than optimizing for prevalence-driven accuracy alone.

Model checkpoints are selected using a composite validation criterion that jointly evaluates discrimination (AUROC), threshold-dependent performance (F1-score and balanced accuracy), and explicit penalties for clinically undesirable error patterns, particularly excess false-negative predictions and large sensitivity–specificity imbalance. This criterion reflects the clinical basis that missed compromises carry greater risk than precautionary escalation, and it ensures that selected models prioritize safety-aligned operating behavior over marginal gains in aggregate accuracy.

Together, these design choices prevent dominance of the majority normal class and ensure that the final model converges on a clinically meaningful operating point, rather than one optimized solely for statistical performance.

Experimental design and statistical analysis

The experimental design follows a predefined analysis plan structured around four components: (i) how the model was developed and evaluated internally and externally; (ii) how performance was assessed at the intended clinical operating point; (iii) how uncertainty and variability were quantified; and (iv) how individual design choices were tested to assess their contribution (ablation analysis).

Validation design

Internal model evaluation was performed using stratified cross-validation on the CTU-UHB dataset. All CTG windows from a given patient were assigned to the same fold, preventing information leakage between training and evaluation. Folds were stratified to ensure that the rare outcome of severe fetal acidemia was proportionally represented, avoiding artificial differences in prevalence across evaluation splits.

Three folds were selected to balance the competing requirements of a sufficiently large training partition (n = 368 records per fold) to support the multi-stage curriculum, and a robust out-of-sample partition (n = 184) for evaluation. The SPaM dataset was withheld entirely from model development and used solely for post-training calibration assessment. Importantly, the decision threshold was not adjusted using SPaM, ensuring that this evaluation reflects transferability rather than re-optimized performance.

Threshold selection

The model produces a continuous probability of severe fetal acidemia, which must be translated into a binary decision to support clinical use (e.g., whether to escalate review). The operating threshold (≈0.52) was determined using the CTU-UHB dataset only, based on cross-validation. Specifically, thresholds were evaluated on out-of-sample predictions to identify a point that balances sensitivity (detection of high-risk cases) and specificity (avoiding excessive false alarms), with priority given to minimizing missed cases of severe acidemia.

Importantly, the threshold was fixed after internal validation and was not adjusted using the external SPaM dataset. This ensures that evaluation on SPaM reflects the model’s true transferability to new clinical settings, rather than performance that has been re-optimized for that dataset.

Performance metrics

Performance was assessed separately for each task to reflect both overall discrimination and clinically relevant operating characteristics.

Binary risk classification (severe fetal acidemia): Performance was assessed using both overall discrimination and threshold-dependent clinical metrics. Overall discrimination was summarized using the area under the receiver operating characteristic curve (AUROC), which reflects the model’s ability to separate high-risk and low-risk cases across all possible thresholds. At the intended operating threshold, sensitivity, specificity, F1-score, balanced accuracy, and Youden’s J were reported. Youden’s J is defined as

These metrics characterize the model’s behavior as a decision-support tool in clinical practice, where a specific threshold is used to guide escalation. Emphasis is placed on maintaining high sensitivity (identifying cases at risk of severe acidemia) while limiting excessive false positives, reflecting the clinical need to reduce missed cases without overburdening care pathways with unnecessary interventions.

Continuous pH regression (severity estimation): Severity estimation performance was evaluated using measures that capture both numerical accuracy and agreement between predicted and observed cord-blood pH values. These metrics are root-mean-square error (RMSE) in pH units and Lin’s concordance correlation coefficient (CCC). CCC quantifies agreement along the line of perfect concordance and penalizes both random error and systematic bias:

where ρ denotes the Pearson correlation coefficient, the respective means, and the standard deviations of predicted and true pH values. This combination ensures that model performance reflects not only whether predictions follow the correct direction of change, but also whether predicted pH values are accurate in magnitude and therefore meaningful for clinical interpretation.

CTG event segmentation (interpretability): For interpretability of model outputs, segmentation performance was evaluated using macro-averaged Dice coefficient and mean Intersection over Union (mIoU). These overlap-based measures assess how well the model identifies clinically relevant signal patterns (e.g., abnormal segments) along the CTG trace. Macro-averaging ensures that rare but clinically important events are weighted equally to more common patterns, preventing performance estimates from being dominated by normal signal segments.

Formal definitions of all metrics are provided in Supplementary Methods S3.

Uncertainty and confidence intervals

All primary metrics are accompanied by 95% confidence intervals, derived using resampling procedures applied to out-of-fold (OOF) predictions. Reporting both average performance and variability across folds provides transparency regarding the stability of results.

For external assessment on SPaM, the evaluation focused on calibration, comparing the model’s predicted abnormality rate at the pre-specified threshold of 0.5 with the cohort’s known prevalence. As the threshold was not re-optimized, this assessment reflects probability consistency rather than a full external re-estimation of sensitivity and specificity.

Ablation and component testing

To assess the contribution of every design choice of CTG-FRAME, a systematic ablation analysis was performed in which one component or training decision was altered at a time while all others were held constant. Each variant was evaluated using the same data splits as the full model, enabling paired comparisons.

Comparisons focused on changes in discrimination performance and clinically relevant error patterns. Statistical testing and effect size estimation were used to distinguish meaningful improvements from random variation, with appropriate correction for multiple comparisons. Statistical comparisons employed a paired t-test on fold-wise AUROC, with Cohen’s d reported as an effect size measure, and McNemar’s test on OOF binary predictions at a fixed threshold of 0.5 . A Bonferroni correction was applied to control the family-wise error rate across multiple comparisons. The ablation scenarios included removal of self-supervised pre-training, alternative strategies for integrating FHR and UC signals, simplified signal preprocessing, replacement of the temporal encoder, and naïve single-stage training without a curriculum. Full numerical results and statistical details are reported in section “Signal context and data integrity” and Supplementary Table S5.

Explainability of the model decision

To support clinical interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) is used to visualize which time periods of the CTG recording most influenced a given model prediction. In practical terms, Grad-CAM identifies segments of the FHR and UC traces that contributed most strongly to the model’s assessment, and displays this information as colour-coded heatmaps overlaid on the original signals. Regions highlighted by Grad-CAM, therefore, indicate when in the trace the model found evidence supporting either a pathological or a reassuring classification, rather than providing an abstract or opaque explanation.

Saliency maps are generated at multiple levels of the model to reflect its hierarchical reasoning process. Early layers correspond to stages of processing that focus on local signal morphology, such as individual decelerations, accelerations, or variability within short time windows. Intermediate layers capture the temporal relationship between FHR and uterine contractions, analogous to clinical assessment of whether decelerations occur early, coincident with, or late relative to contractions. Deeper layers integrate information across longer time spans, capturing global temporal context across labor, such as whether abnormal patterns are isolated, repetitive, or progressively worsening. All saliency maps are time-aligned to the original CTG traces, enabling clinicians to directly relate model explanations to familiar signal features observed during bedside interpretation.

Because saliency values are not directly comparable across different network layers, each layer represents information at a different level of abstraction, and interpretation is performed in aggregate. Rather than relying on isolated activations at a single layer, attention is directed toward consistent patterns that recur across multiple layers and across both signal modalities. This approach reduces the risk of over-interpreting transient or spurious activations that may arise from scaling effects or non-linear transformations within the model’s network.

Explanations are always considered alongside quantitative performance metrics and are intended to support clinical review, not to replace clinical judgment. Specifically, Grad-CAM is used to verify that model predictions are grounded in clinically meaningful signal regions, such as late or prolonged decelerations, sustained low variability, or reassuring accelerations with prompt post-contraction recovery, rather than artifactual noise or signal dropout. Avoiding over-interpretation of attribution maps means recognizing that saliency indicates contributing evidence, not causality, and that isolated highlighted regions should not be treated as definitive diagnostic findings in the absence of broader contextual consistency.

This structured, multi-level explainability strategy ensures that CTG-FRAME’s outputs remain auditable, physiologically plausible, and aligned with established clinical reasoning, while mitigating common pitfalls of attribution methods, including sensitivity to local scaling, layer-specific artifacts, and false certainty derived from single-map inspection.

Results

CTG-FRAME was evaluated across four sequential analyses. First, internal performance was established on the CTU-UHB dataset using cross-validation for severe fetal acidemia classification, continuous pH severity estimation, and event localization. Second, an ablation study quantified the contribution of each architectural and training component. Third, qualitative explainability analyses using saliency mapping were conducted to verify that predictions were grounded in physiologically meaningful signal patterns. Finally, an external calibration assessment on the SPaM cohort examined cross-site transferability without retraining the model.

Internal validation

Table 2 specifies the internal validation performance across folds and the pooled out-of-fold results.

Clinical operating point (pooled out-of-fold)

At the intended clinical operating threshold (≈0.52), the pooled out-of-fold predictions yield a conservative, escalation-consistent performance profile. Discrimination profile was high, AUROC = 0.888 (95% CI 0.832–0.937), sensitivity = 0.850 (0.771–0.909), specificity = 0.964 (0.950–0.986), balanced accuracy = 0.907, and Youden’s J = 0.814. This balance reflects a deliberate asymmetry appropriate for intrapartum escalation decisions, in which the clinical cost of undetected acidemia substantially exceeds that of a false-positive referral for review.

Per-fold stability versus pooled performance

Across individual folds, discrimination performance was consistently high (mean AUROC = 0.974 ± 0.013). The lower pooled AUROC reflects the application of a single operating threshold across heterogeneous patient subgroups, rather than variation in underlying discrimination. This distinction highlights the interpretive difference between within-fold performance estimates and pooled, deployment-relevant summaries, and suggests that modest threshold adjustment may be required across subpopulations in real-world settings.

pH severity regression (pooled out-of-fold)

The regression head produced numerically accurate and concordant severity estimates, with RMSE = 0.563 pH units (95% CI 0.516–0.612) and concordance correlation coefficient (CCC) = 0.823 (0.780–0.856). This indicates both low average error and limited systematic bias, supporting the use of model outputs as a continuous severity marker rather than a binary surrogate alone. These results are consistent with the multi-task design, in which classification and event-localization objectives regularize the shared representation.

Event segmentation (interpretability)

Despite the rarity and inherent ambiguity of event-level annotations, the model demonstrates meaningful localization of clinically informative signal regions. Per-fold estimates yield macro-Dice = 0.458 ± 0.089 (macro-mIoU = 0.395 ± 0.109; Fold 3 Dice = 0.583), while the pooled analysis yields macro-Dice = 0.333 (95% CI 0.269–0.395) and macro-mIoU = 0.277 (0.238–0.327). While absolute values are moderate, as expected given annotation uncertainty, the segmentation head is not intended to replace clinical labeling. Instead, it identifies temporal regions contributing evidence to the case-level prediction, enabling clinicians to audit why a trace was flagged rather than relying on a black-box output.

Ablation result

A comprehensive ablation study was conducted to validate each architectural and pre-processing decision within CTG-FRAME (Table 3). Beginning from the complete multi-task model as the reference configuration, individual components were systematically removed or replaced, and the resulting models were re-evaluated on identical cross-validation splits. This design isolates the contribution of each element and demonstrates that optimal performance arises from the coordinated integration of multiple components rather than from any single module in isolation. Full statistics and supplementary comparisons are provided in Supplementary Table S4.

Table 3.

Statistical ablation study results comparing component ablations to the baseline

Component	AUROC	Composite Score	F1 Score	Balanced Acc	Sensitivity	Specificity	pH RMSE	pH CCC	Dice (Macro)
CTG-FRAME (Proposed)	0.888 [0.832–0.937]	0.536 [0.483–0.578]	0.853 [0.800–0.909]	0.907 [0.868–0.940]	0.850 [0.771–0.909]	0.964 [0.950–0.986]	0.563 [0.516–0.612]	0.823 [0.780–0.856]	0.333 [0.269–0.395]
Replace Cross-Attention with Concat	0.805 [0.746–0.857]	0.416 [0.362–0.467]	0.637 [0.575–0.716]	0.798 [0.753–0.841]	0.752 [0.664–0.827]	0.843 [0.810–0.910]	0.763 [0.711–0.812]	0.616 [0.551–0.676]	0.294 [0.252–0.340]
Replace Cross-Attention with Add	0.842 [0.789–0.898]	0.482 [0.432–0.528]	0.771 [0.706–0.833]	0.861 [0.823–0.899]	0.788 [0.712–0.863]	0.934 [0.893–0.973]	0.687 [0.637–0.737]	0.699 [0.639–0.748]	0.341 [0.276–0.401]
Replace ResNet-TCN with Standard CNN	0.849 [0.788–0.902]	0.496 [0.448–0.541]	0.782 [0.729–0.845]	0.875 [0.838–0.911]	0.823 [0.748–0.885]	0.927 [0.906–0.959]	0.701 [0.629–0.774]	0.698 [0.632–0.757]	0.262 [0.233–0.310]
+ Hand-Crafted Features	0.662 [0.614–0.712]	0.260 [0.202–0.315]	0.458 [0.379–0.528]	0.662 [0.614–0.712]	0.504 [0.415–0.600]	0.820 [0.784–0.856]	0.987 [0.920–1.055]	0.058 [0.031–0.084]	0.000 [0.249–0.250]
-MoCo Pre-training	0.768 [0.720–0.813]	0.365 [0.319–0.406]	0.540 [0.459–0.602]	0.738 [0.694–0.785]	0.717 [0.680–0.852]	0.759 [0.639–0.791]	1.067 [1.003–1.127]	0.263 [0.192–0.330]	0.247 [0.246–0.248]
Advanced to Simplified Signal Cleaning	0.792 [0.731–0.855]	0.422 [0.369–0.479]	0.709 [0.633–0.783]	0.812 [0.772–0.860]	0.690 [0.619–0.791]	0.934 [0.878–0.967]	0.765 [0.701–0.832]	0.597 [0.522–0.667]	0.267 [0.242–0.298]

Baseline (full model)

The complete integrated model achieved AUROC = 0.888 [0.832–0.937], pH RMSE = 0.563 [0.516–0.612], and macro-Dice = 0.333 [0.269–0.395]. These values constitute the reference benchmark for all ablation comparisons and reflect the clinically aligned operating point reported in section “Primary outcome and reference standard”.

MoCo v2 pre-training

Removing self-supervised pre-training (replacing it with random encoder initialization) produced a marked decline in classification performance (AUROC = 0.768), supported by strong statistical evidence (t = 11.36, p = 0.0077; McNemar’s , p < 0.0001). The effect on pH regression was more severe: RMSE nearly doubled to 1.067 [1.003–1.127]. Collectively, these findings indicate that MoCo v2 pre-training establishes the physiological feature representations upon which the severity regression head depends; in its absence, the model fails to generalize from sparse supervision to accurate continuous pH estimation.

Hand-crafted features

The inclusion of 42 engineered CTG features and associated metadata as auxiliary inputs proved unequivocally detrimental, yielding AUROC = 0.662 with highly significant performance discordance (McNemar’s p < 0.0001). Both auxiliary output heads failed substantially (pH RMSE = 0.987; macro-Dice = 0.000), indicating that the introduction of coarse, pre-defined feature summaries disrupts the fine-grained, end-to-end representations acquired by the encoders, generating conflicting gradient signals and diminishing multi-task synergy.

Cross-attention fusion

Replacing bidirectional FHR–UC cross-attention with simple concatenation fusion substantially reduced classification performance (AUROC = 0.805; McNemar’s , p < 0.0001; Cohen’s d = 2.474). An additive fusion strategy yielded a modest improvement but remained significantly inferior to the full model (AUROC = 0.842; ; p = 0.0429). These comparisons confirm that explicitly modeling inter-signal temporal dependencies—rather than merely co-representing FHR and UC—is essential for detecting clinically decisive patterns, such as the latency of decelerations relative to uterine contractions.

Backbone and pre-processing

Replacing the ResNet-TCN stem with a standard convolutional network produced a notable reduction in segmentation performance (macro-Dice: 0.333 → 0.262), supporting the use of the hybrid backbone architecture in this multi-task context. Substituting the full signal cleaning pipeline with a simplified preprocessing protocol similarly degraded classification performance (AUROC = 0.792; McNemar’s , p = 0.0052), underscoring that high-quality, standardized signal inputs are a prerequisite for reliable model outputs.

Synthesis

The ablation study reveals a clear performance hierarchy and, more critically, demonstrates that state-of-the-art performance is contingent upon the joint contribution of four components: (1) self-supervised pre-training via MoCo v2, (2) end-to-end representation learning without hand-crafted feature inputs, (3) bidirectional cross-attention fusion for temporally-aware FHR–UC interaction, and (4) a domain-specific hybrid backbone paired with a validated signal preprocessing pipeline. The removal of any individual component degrades performance on one or more tasks, confirming that the final model’s capabilities emerge from the principled integration of these elements rather than from any single design decision.

Analysis of the difficulty-aware curriculum training result

The difficulty-aware, multi-stage curriculum constitutes a core enabling mechanism of CTG-FRAME. To quantify its contribution, performance was compared against a single-stage baseline trained end-to-end on the full, imbalanced dataset. The contrast is decisive: the single-stage model failed to converge to a clinically useful solution (AUROC ≈ 0.500; specificity = 0.000; accuracy = 0.065), defaulting to near-universal positive-class prediction, a canonical failure mode under severe class imbalance.

The curriculum proceeds through three sequential stages (W0 → W1 → W2), each serving a distinct learning objective and contributing measurable, incremental improvements to model performance (see Table 4).

Table 4.

Impact of curriculum learning stages (mean ± SD across folds)

Task	Metric	Naive Single-Stage Training	Difficulty-Aware Multi-Stage Training
			W0	W1	W2
Binary Fetal Acidemia	AUROC	0.500	0.687 (±0.079)	0.741 (±0.104)	0.974 (±0.016)
	Sensitivity (Recall)	1.000	0.851 (±0.177)	0.279 (±0.483)	0.930 (±0.036)
	Specificity	0	0.500 (±0.306)	0.897 (±0.179)	0.932 (±0.035)
	F1-Score	0.122	0.718 (±0.041)	0.279 (±0.483)	0.930 (±0.036)
	Youden’s J Statistic	0	0.351 (±0.161)	0.175 (±0.304)	0.862 (±0.071)
	Composite Score	0	0.351 (±0.074)	0.151 (±0.282)	0.599 (±0.026)
	Accuracy	0.065	0.672 (±0.084)	0.886 (±0.086)	0.931 (±0.035)
	Balanced Accuracy	0.500	0.675 (±0.080)	0.588 (±0.152)	0.931 (±0.035)
	Threshold	0.100	0.3622 (±0.0945)	0.7813 (±0.2916)	0.5559 (±0.0487)
Continuous pH Regression	RMSE (pH units)	0.105	1.122 (±0.293)	0.884 (±0.039)	0.476 (±0.141)
	Concordance Corr. Coeff. (CCC)	0.002	0.291 (±0.255)	0.293 (±0.301)	0.904 (±0.055)
CTG Event Segmentation	Dice Score (Macro Avg.)	0.249	0.240 (±0.015)	0.239 (±0.018)	0.458 (±0.109)
	mIoU (Macro Avg.)	0.240	0.231 (±0.029)	0.230 (±0.031)	0.395 (±0.134)

W0: warm-up (baseline)

W0 establishes an initial, non-random decision boundary: AUROC = 0.687 ± 0.079, with high sensitivity (0.851) but low specificity (0.500). This positive-class bias is an expected consequence of the warm-up configuration and is intentionally accepted as the foundation upon which subsequent stages build.

W1: mid-point (bias inversion)

W1 introduces a deliberate corrective rebalancing. By restructuring the sampling distribution toward hard negatives and minority-class instances, specificity increases markedly to 0.897, even as sensitivity temporarily decreases to 0.279 and composite metrics soften. This stage compels the model to acquire discriminative representations of normal physiological patterns, a prerequisite for achieving balanced performance in the final stage.

W2: final synthesis (full fine-tuning and hard-mining)

W2 yields the performance balance: AUROC = 0.974 ± 0.016, sensitivity = 0.930, specificity = 0.932. Youden’s J index advances from approximately 0.35 at W0 to 0.814 at W2, with balanced accuracy reaching 0.907. W2 thus integrates the broad detection sensitivity established in W0 with the specificity-oriented discriminative capacity acquired in W1, converging on a high-performing, clinically balanced operating region.

Benefits extend beyond classification

The advantages of difficulty-aware curriculum training extend coherently across all task heads. During the early and intermediate stages (W0–W1), pH severity estimation remains limited (CCC ≈ 0.29), reflecting incomplete alignment between learned signal representations and continuous acid–base outcomes. In W2, concordance increases substantially (CCC = 0.904), accompanied by a reduction in RMSE of more than 57%, indicating both improved accuracy and reduced systematic bias. Event localization performance exhibits a similar trajectory, increasing from a low baseline (macro-Dice ≈ 0.24) to approximately double that value by W2. These cross-task improvements indicate that the shared representations learned through staged training become progressively richer, physiologically coherent, and more generalizable than those obtained under single-stage optimization.

In the absence of curriculum learning, extreme class imbalance drives the model toward a degenerate result characterized by near-universal positive prediction. By contrast, the curriculum induces a structured learning trajectory: an initial, sensitivity-biased configuration that prioritizes detection (W0); a corrective phase that consolidates representations of normal physiology and restores specificity (W1); and a final stage in which these complementary properties are reconciled into a balanced, deployment-consistent operating regime (W2). The consistent improvement across stages, together with concurrent gains in severity estimation and event localization, demonstrates that difficulty-aware curriculum training is not merely beneficial but necessary for stable, clinically meaningful learning in this rare-event, multi-task intrapartum setting.

Explainability analysis

Saliency analysis confirms that CTG-FRAME’s predictions are grounded in clinically plausible signal evidence. Across successive network depths, attribution follows a hierarchical pattern that closely parallels established CTG interpretation practice: identification of local morphological features, contextual interpretation relative to uterine contractions, assessment of temporal persistence and clustering, and synthesis into a global risk judgment.

Level 1: encoder-level morphology detection (local pattern recognition)

At the ResNet–TCN encoder stage, saliency maps highlight local morphological features within a short temporal context, including FHR nadirs during decelerations, contraction peaks and relaxation slopes, and short-term variability characteristics. This corresponds to the initial step in clinical trace assessment, in which individual signal segments are inspected for abnormal morphology independent of broader context.

Level 2: cross-attention timing resolution (FHR–UC temporal relationship)

At the bidirectional FHR–UC cross-attention module, saliency emphasizes inter-signal timing relationships, enabling assessment of whether FHR changes occur before, during, or after UC. Windows in which deceleration onset follows contraction onset, consistent with late deceleration patterns, receive elevated FHR-to-UC attribution, whereas early or variable decelerations produce more diffuse attention. This behavior aligns with the established clinical importance of deceleration latency as an indicator of uteroplacental insufficiency and hypoxic stress.

Level 3: transformer self-attention window selection (temporal prioritization across labor)

Within the Transformer encoder, self-attention selectively prioritizes the most informative temporal windows for the case-level decision. Elevated attribution tends to concentrate in late labor and around repeated abnormal episodes, reflecting the clinical principle that escalation decisions are driven by persistence, accumulation, and temporal clustering of abnormalities rather than by isolated events.

Level 4: evidence synthesis across task heads (case-level convergence)

At the final stage, the classification, pH regression, and event-segmentation heads converge on consistent evidential regions. High predicted risk coincides with low estimated pH and low uncertainty in unambiguous cases, while the segmentation head localizes the same temporal windows emphasized at earlier stages. This cross-task agreement strengthens the face validity of the model’s outputs and provides clinicians with an auditable basis for understanding why a case is flagged.

Case-level analysis: illustrative examples

In correctly classified acidemia cases (Table 5), saliency progressively narrows and concentrates with increasing network depth, reflecting a transition from broad pattern scanning to focused identification of clinically decisive events.

Table 5.

Hierarchical attention refinement in correctly classified acidemia cases

Across the cases, attribution consistently emphasizes late and prolonged decelerations as well as sustained low variability, aligning with established markers of intrapartum compromise. In a representative case featuring a deep, prolonged deceleration (Record #1017), encoder-level saliency is initially diffuse, scanning broadly across the trace to identify potentially abnormal regions. At the cross-attention stage, attribution becomes concentrated on the window in which the FHR nadir follows the peak of the UC, consistent with a canonical late deceleration pattern. The Transformer encoder subsequently isolates this episode as the dominant contributor to the case-level decision, and the final classification layer preserves this focus through to the “Acidemia” prediction.

A complementary pattern is observed in a case characterized by persistently reduced variability (Record #1479). In this case, early-stage attribution highlights extended flat segments of the trace, while deeper self-attention layers progressively concentrate on the longest and least variable interval, which is ultimately carried through to the final classification. This evolution mirrors clinical reasoning in which concern increases when low variability is both sustained and temporally dominant.

In correctly classified normal cases (Table 6), attribution instead emphasizes reassuring physiological features, including accelerations, prompt post-contraction recovery, and stable variability. In a trace with clear accelerations (Record #1317), encoder-level saliency highlights acceleration morphology, the cross-attention module confirms timely heart-rate recovery relative to contractions, and the Transformer prioritizes a prototypical acceleration window as primary evidence supporting a “Normal” classification. In a more complex but non-pathological trace with high variability (Record #1313), self-attention remains distributed across multiple segments without converging on a single dominant abnormal episode, a pattern consistent with the absence of fetal acidemia rather than masked pathology.

Table 6.

Validation of reassuring pattern identification in correctly classified normal cases

This table shows the model’s ability to classify non-acidemia cases by focusing on reassuring features. The top row, the Aggregated Saliency Map, shows the overall signal importance from the initial FHR encoder. The bottom rows display attention from progressively deeper layers for a classic normal trace (Record #1137) and a more complex normal trace (Record #1313). For the normal case, this map correctly highlights regions with reassuring, which the raw contribution map (bottom panel) shows as red for normal and blue for an acidemia prediction

Grad-CAM attributions are computed at each network layer and temporally aligned to the original CTG signals, enabling inspection of the precise loci of evidential emphasis within each trace. Because attention and feed-forward layers transform information via different mechanisms, saliency intensity is not directly comparable across layers. Accordingly, attribution maps are interpreted in aggregate, across multiple layers and both signal modalities, and always in conjunction with quantitative performance metrics. This approach mitigates common attribution pitfalls, including single-layer artifacts, transient activations, and false certainty derived from isolated highlights.

Across all illustrative cases, CTG-FRAME integrates local morphological features, FHR–UC timing relationships, and global temporal context in a manner consistent with established intrapartum reasoning. Cross-task agreement between classification, severity estimation, and event localization further strengthens the face validity of the model’s outputs and supports their use as auditable clinical decision support, rather than autonomous diagnosis.

This table demonstrates the model’s reasoning in identifying pathological patterns. Each column is a unique patient record correctly classified with acidemia. The top row, the Aggregated Saliency Map, visualizes the overall importance derived from the initial FHR feature encoder. This map highlights the specific windows that contributed most significantly to the final “Acidemia” prediction before being contextualized by the deep Transformer layers. The bottom rows show Grad-CAM heatmaps from progressively deeper layers, illustrating how the model refines its focus from broad features (Initial Fusion) to specific events within the Transformer. The red overlay indicates regions of high model importance, while the bottom panel displays the raw signal contribution (red: excitatory for acidemia, blue: suppressive).

External calibration assessment

As an assessment of cross-site generalizability in the absence of retraining or threshold optimization, CTG-FRAME was applied to the SPaM cohort (n = 300; multi-centre) and evaluated for calibration rather than discrimination. Given a known severe acidemia prevalence of approximately 20% in this cohort, a well-calibrated model would be expected to predict an abnormal rate within that range when evaluated at a standard operating threshold.

At the default decision threshold (p = 0.5), CTG-FRAME classified 72 of 300 cases (24.0%) as abnormal, closely approximating the cohort prevalence and indicating that output probabilities are reasonably calibrated to the base rate without site-specific tuning. Upon adjustment to a threshold of p = 0.523, the predicted abnormal rate aligned precisely with the 20.0% cohort prevalence, suggesting that the model’s probability estimates are well-ordered across cohorts and that rate alignment is achievable through a minor operating-point adjustment.

This analysis constitutes a calibration assessment rather than a full external validation of sensitivity and specificity: the SPaM cohort was employed exclusively for post-training calibration and was not used for model fitting or threshold selection. These findings provide preliminary evidence that CTG-FRAME’s learned representations capture physiological patterns that transfer across clinical sites; however, prospective, multi-centre studies with confirmed outcome labels remain necessary to establish clinical generalizability and to re-estimate operating characteristics in independent settings.

Comparison with state-of-the-art methods

Table 7 situates CTG-FRAME within the current landscape of automated CTG analysis methods evaluated on the CTU-UHB dataset, encompassing approaches that prioritize lightweight deployment, multi-modal fusion, or narrowly scoped binary classification across varying pH thresholds. The substantial variation in reported performance underscores that headline metrics, particularly AUROC, are strongly influenced by task formulation and training strategy, and cannot be attributed to architectural differences alone.

Table 7.

Comparison of CTG-FRAME with related works in automated CTG analysis (trained and tested on CTU-UHB dataset)

Approach	Key Differentiators and Rationale	pH Threshold	Key Reported Metric(s)
Task: Severe Acidemia Classification (pH ≤ 7.05)
LW-FHRNet (ShuffleNet + ECA Attention) [32]	Focus on a lightweight (0.33 M params), deployable model using channel-attention for high efficiency.	<7.05	Acc: 95.2%, Se: 90.5%, Sp: 100.0%, F1: 0.95, AUROC: 0.9796
MENet (Multi-modal) [33]	Multi-modal framework designed to fuse different data types for a holistic analysis.	<7.05	Acc: 94.0%, Se: 94.7%, Sp: 93.4%, F1: 0.94
DeepCTG1.5 [17]	A multimodal approach that explicitly combines learned CTG features with clinical data.	<7.05	AUROC: 0.71
CTG-FRAME (This Work)	+ Comprehensive Multi-Stage Curriculum: Explicitly handles class imbalance and hard examples. + Self-Supervised Pre-training (MoCo v2): Learns robust features from unlabeled data. + Advanced Bidirectional Fusion: Synergistically models the critical FHR-UC relationship. + Multi-Task and Interpretable Outputs: Provides classification, regression, and segmentation with Grad-CAM for clinical trust.	≤7.05	AUROC: 0.888, Se: 0.850, Sp: 0.964, F1: 0.853 Acc: 0.940, Bal. Acc: 0.907, Youden’s J: 0.813
Task: Mild/ Moderate Acidemia Classification (pH ≤ 7.15)
Custom CNN on CWT Spectrogram [34]	Converts 1D signal to 2D CWT spectrogram for image-based analysis. Computationally expensive.	<7.15	Acc: 98.3%, Se: 98.2%, Sp: 94.9%, AUROC: 0.9782
CNN + BiLSTM + Attention [35]	Hybrid CNN-RNN using wavelet features and an attention mechanism to detect subtle decelerations.	≤7.15	Acc: 75.2%, Se: 70.8%, QI: 72.29%
MENet (Multi-modal) [33]	Multi-modal framework designed to fuse different data types for a holistic analysis.	<7.15	Acc: 86.3%, Se: 93.3%, Sp: 81.9%, F1: 0.87
Task: Other or Unspecified Thresholds
1D CNN on Time-Series Segments [36]	Segment-level CNN analysis designed to detect abnormalities earlier than full-trace methods.	<7.2	Acc: 80.0%, Se: 79.0%, AUROC: 0.86
CNN and Voting Mechanism [37]	Ensemble approach using multiple CNNs with a voting mechanism on the last 30 mins of the trace.	Not Specified	Set 1: Acc: 89.3%, Se: 73.8%, Sp: 90.7% <br>Set 2: Acc: 80.0%, Se: 80.9%, Sp: 79.9%
DeepCTG, Logistic Regression [38]	Combines deep learning features (DeepCTG) with a traditional logistic regression classifier.	Not Specified	AUROC: 0.74
Auxillary Tasks (This Work)
CTG-FRAME (This Work)	(As above)	pH Regression	RMSE: 0.563, CCC: 0.823
CTG-FRAME (This Work)	(As above)	Segmentation	Dice: 0.333, mIoU: 0.200

The majority of state-of-the-art methods reporting high AUROC values, including LW-FHRNet and MENet, address single-task binary classification of severe acidemia (pH ≤ 7.05). These models are typically optimized to maximize class separability under a fixed threshold and do not attempt severity estimation, evidence localization, or uncertainty quantification. Consequently, while such models may achieve strong discrimination on the CTU-UHB dataset, they offer limited support for clinical verification and escalation decision-making, in which the interpretive basis for a classification is as clinically consequential as the classification itself.

Among the closest-performing models evaluated at the same pH threshold, CTG-FRAME’s AUROC (0.888) is lower than that of the strongest single-task classifiers; however, this difference must be interpreted in context. CTG-FRAME is designed and evaluated as a multi-task clinical decision-support system rather than as a pure classifier. Its optimization explicitly accepts a modest reduction in threshold-free discrimination in exchange for gains in balanced operating behavior, severity estimation, and interpretability, as reflected by high specificity (0.964), balanced accuracy (0.907), and Youden’s J (0.813) at the intended clinical operating point.

A further source of performance variation across Table 7 is the choice of pH threshold. Models targeting mild or moderate acidemia (pH ≤ 7.15) or higher thresholds report substantially higher accuracy and AUROC values, reflecting the greater prevalence and increased class separability associated with less severe endpoints. CTG-FRAME deliberately targets the rarer, higher-risk endpoint (pH ≤ 7.05), where class imbalance is more pronounced and the clinical cost of false-negative predictions is substantially greater. Cross-category performance comparisons are therefore not directly interpretable without accounting for the differential difficulty imposed by threshold selection.

CTG-FRAME further distinguishes itself from prior work in its approach to class imbalance. Rather than relying solely on loss re-weighting or resampling strategies, it integrates self-supervised pre-training, focal loss, and a difficulty-aware multi-stage curriculum that progressively restructures the training distribution. Ablation analysis confirms that removal of the curriculum results in collapse to majority-class prediction, indicating that CTG-FRAME’s performance reflects a stable operating regime under severe class imbalance rather than threshold optimization.

CTG-FRAME is the only method in Table 7 to jointly report continuous pH regression (RMSE = 0.563, CCC = 0.823) and temporal event localization (macro-Dice = 0.333), providing clinically interpretable evidence alongside the primary classification output. Critically, these auxiliary outputs are not post-hoc additions; they function as regularizers that stabilize representation learning and improve generalization, particularly under rare-event training conditions.

In summary, the result comparison illustrates a fundamental trade-off in automated CTG analysis: while several state-of-the-art models maximize AUROC for narrowly defined classification tasks, CTG-FRAME prioritizes clinically aligned decision support by balancing discrimination with calibration, severity estimation, interpretability, and robustness under extreme class imbalance. This positions CTG-FRAME not as the highest-AUROC classifier in the comparison, but as the most comprehensive and translationally oriented CTG decision-support framework currently reported on this benchmark. These comparative findings provide the interpretive context for the Discussion, in which CTG-FRAME’s performance trade-offs are examined in relation to clinical risk, deployment constraints, and the requirements of trustworthy intrapartum decision support.

Discussion

This section interprets the principal findings in the context of related work, articulates the methodological and clinical implications, and concludes with a transparent account of study limitations and future research directions. This section focuses on why CTG-FRAME performs as observed, which design decisions are most consequential, and how these decisions align with the demands of real-world intrapartum clinical decision-making.

Interpretation of key findings

The performance of CTG-FRAME emerges from the coordinated interaction of multiple design decisions rather than from any single architectural component. As summarized in Table 7, headline discrimination metrics vary substantially across studies evaluated on the CTU-UHB dataset, reflecting differences in task definition, class imbalance handling, and clinical scope. CTG-FRAME is distinguished not by maximizing AUROC in isolation, but by achieving a methodologically validated, clinically balanced operating point while simultaneously satisfying multi-task clinical requirements.

CTG-FRAME estimates the risk of severe fetal acidemia using umbilical arterial cord pH as an objectively measured but partial proxy of fetal compromise. Umbilical arterial pH is a valid, objective, and reproducible reference standard for severe fetal acidemia, reflecting systemic acid–base status at birth. However, it is not a direct measure of tissue-level hypoxia, cellular metabolic acidosis, or long-term neurological injury. Accordingly, while cord blood pH provides a clinically meaningful indicator of acute intrapartum acidemia, it does not capture the broader spectrum of fetal oxygenation dynamics or downstream outcomes. CTG-FRAME is therefore designed to support intrapartum escalation decisions associated with acute acidemia risk, rather than to diagnose fetal hypoxia or predict long-term neonatal morbidity.

Robustness to incomplete signal is a critical consideration in intrapartum monitoring, where transient signal loss is common. Sensitivity analysis demonstrated that CTG-FRAME maintains discriminative performance across shorter reconstruction thresholds (15, 30, and 45 seconds), indicating that model performance is not dependent on a specific preprocessing parameter. However, shorter thresholds were associated with increased variability and reduced agreement in continuous pH estimation, consistent with fragmentation of temporally coherent physiological patterns. These findings suggest that moderate reconstruction supports the preservation of meaningful temporal structure, rather than restoring unobserved physiology. Accordingly, the selected 60-second threshold represents a pragmatic balance between robustness to artifact and preservation of clinical interpretability. Importantly, reconstructed segments are treated as regions of uncertainty rather than confirmed signal, and model outputs should be interpreted in the context of signal quality and availability.

Training curriculum as a prerequisite for convergence

The most foundational finding is that the difficulty-aware multi-stage curriculum functions as an enabling prerequisite rather than an incremental optimization. Single-stage training collapses to majority-class prediction (AUROC ≈ 0.50; specificity ≈0.000), a well-documented failure mode in severely imbalanced biomedical datasets [30, 39]. This outcome confirms that, for rare but high-consequence endpoints such as severe acidemia, conventional training pipelines provide insufficient gradient signal to develop meaningful minority-class representations.

The curriculum addresses this through a purposeful, non-linear progression across three stages. Stage W0 establishes an intentionally biased initial configuration with high sensitivity and low specificity, consistent with the expected early-stage failure mode. Stage W1 corrects this bias by restructuring the training distribution toward hard negatives, compelling the model to acquire discriminative representations of normal physiology—a counterintuitive but clinically essential step, evidenced by specificity increasing from 0.500 to 0.897. Stage W2 consolidates these representations through full fine-tuning, yielding a balanced and clinically viable model, with Youden’s J advancing from 0.351 at W0 to 0.862 at W2 and balanced accuracy reaching 0.907.

These findings extend curriculum learning theory beyond standard easy-to-hard scheduling [40], demonstrating that under conditions of extreme class imbalance, bias-inverting curricula are necessary to preclude degenerate solutions. Compared with data-level remedies such as SMOTE [41, 42]—which risk generating physiologically implausible synthetic samples—this training-level strategy offers a more principled and generalizable approach to rare-event detection in clinical time-series.

Synergistic architecture: ablation-justified complexity

The ablation study provides data-driven justification for each architectural element and demonstrates that peak performance arises from their coordinated synergy rather than from redundant contributions. Removal of MoCo v2 self-supervised pre-training produced near-catastrophic degradation of pH regression performance (RMSE +89%), establishing that contrastive representation learning is a necessary rather than optional component for severity estimation under label scarcity. These findings extend the applicability of self-supervised learning from computer vision to physiological time-series domains, where labeled data are costly to acquire and inherently sparse [43, 44].

Replacing bidirectional FHR-UC cross-attention with simple fusion substantially reduced classification performance, empirically confirming that clinically decisive information resides in inter-signal temporal relationships rather than in mere co-representation [32, 33]. Most notably, the inclusion of hand-crafted CTG features produced a complete performance collapse, contradicting the prevailing assumption in biomedical informatics that engineered features provide additive value to learned representations [45]. For complex, multi-modal physiological signals, these results indicate that end-to-end learned representations and manual feature engineering are not complementary but mutually incompatible.

Consistent with the comparison presented in Table 7, the published literature remains dominated by single-task models [17, 32, 36, 38, 46, 47]. To the authors’ knowledge, CTG-FRAME is among the first CTG frameworks to jointly report severe fetal acidemia classification, continuous pH estimation with uncertainty, and temporal evidence localization within a single end-to-end architecture.

Holistic framework for clinically relevant performance

CTG-FRAME reframes CTG analysis from isolated binary classification to a decision-analytic framework organized around three clinically grounded questions: whether a problem is present (classification), how severe it is (pH regression), and where the supporting evidence is located within the signal (temporal localization). This structure mirrors the sequential reasoning process employed during intrapartum surveillance and contrasts with the binary, single-output design of most existing approaches [48–50].

The selected operating point (sensitivity = 0.850; specificity = 0.964) reflects a harm-asymmetric optimization appropriate for escalation support, in which undetected acidemia carries substantially greater clinical consequence than a false-positive referral for review, yet excessive false-alarm rates would risk eroding clinician trust and system adoption. Composite metrics such as Youden’s J (0.814) and balanced accuracy (0.907) are therefore more informative than AUROC alone for characterizing performance in this setting [63–65]. The reported AUROC of 0.888 should accordingly be interpreted as a conservative, reproducible discrimination estimate achieved under stringent class imbalance constraints, rather than as an optimistic, threshold-optimized result.

Implications

From a computational perspective, CTG-FRAME offers a validated design paradigm for imbalanced, multi-modal, label-scarce physiological time-series data. The difficulty-aware curriculum provides a generalizable training-level solution for rare-event detection in settings where data-level augmentation is insufficient. The demonstrated utility of self-supervised pre-training supports the broader development of foundational models for physiological signals, including ECG, EEG, and photoplethysmography [51, 52]. The bidirectional cross-attention fusion architecture provides a reusable template for modeling inter-signal temporal relationships in other coupled biosignal domains.

From a clinical perspective, CTG-FRAME is most appropriately characterized as an objective decision-support adjunct rather than an autonomous classifier. By jointly providing risk stratification, severity estimation, and temporally localized interpretable evidence, the system is designed to support clinical verification and escalation decisions rather than to supplant clinician judgment. This design philosophy, grounded in transparency through evidence rather than prediction alone, aligns with emerging frameworks for trustworthy clinical AI and with qualitative evidence regarding obstetricians’ expectations for AI-assisted intrapartum support [48, 49]. CTG-FRAME is intended to facilitate clinican review and more informed surveillance, consistent with the role of decision-support technologies in reducing diagnostic uncertainty in time-critical clinical settings.

Scope of inference

The scope of inference of this study is bounded by the characteristics of the CTU-UHB dataset. CTG recordings are limited to the final 90 minutes preceding delivery, with second-stage labour duration restricted to less than 30 minutes. As a result, earlier intrapartum signal evolution is not observed. Some pathological CTG patterns, such as evolving baseline abnormalities or repetitive decelerations, may have emerged earlier and prompted clinical intervention before the recorded window. Consequently, model performance should be interpreted in the context of acute intrapartum assessment rather than continuous longitudinal monitoring, and potential intervention bias cannot be excluded. In addition, restricting the analysis to recordings with a short second stage of labor (<30 minutes) introduces a selection bias toward cases with relatively rapid delivery. This criterion may under-represent prolonged second-stage labors, during which CTG interpretation and decision-making dynamics can differ substantially. Therefore, findings may not generalize to all intrapartum contexts, particularly those involving extended second-stage management.

UC signals in CTU-UHB and SPaM do not uniformly specify whether the measurement was obtained via external tocodynamometry or an intrauterine pressure catheter. As external tocodynamometry reflects relative abdominal wall tension rather than true intrauterine pressure, UC amplitude cannot be interpreted as an absolute measure of contraction strength in mmHg. Accordingly, CTG-FRAME treats uterine activity as a trace-relative temporal signal, emphasizing contraction timing and patterning rather than absolute magnitude. While this improves robustness across heterogeneous monitoring conditions, it limits the physiological interpretation of contraction intensity.

Strengths, limitations, and future research directions

The key strength of this study is its end-to-end, ablation-validated design, which demonstrates that difficulty-aware curriculum training and multi-task learning are necessary conditions for avoiding degenerate solutions in rare-event CTG analysis.

Several limitations warrant acknowledgment. First, the study uses umbilical arterial cord blood pH as the sole outcome, which reflects systemic acidemia at birth and does not directly capture fetal oxygenation dynamics, organ-specific injury, or postnatal adaptation. Clinical outcomes such as Apgar scores, neonatal resuscitation, and longer-term neurodevelopmental impairment arise from more complex pathways and were not available for model development. Accordingly, the present model should be interpreted as supporting intrapartum acidemia risk stratification, rather than comprehensive neonatal prognosis.

Second, the CTU-UHB dataset does not include long-term outcomes, such as cerebral palsy or neurodevelopmental impairment. As these outcomes reflect downstream processes beyond acute intrapartum events, they cannot be inferred directly from CTG signals or cord pH alone, and limit the scope of interpretation. In addition, publicly available CTG datasets contain limited demographic and clinical covariate information, which restricts formal fairness assessment across key subgroups such as ethnicity, socioeconomic status, and comorbidity.

To complement the results, the clinical characteristics of false-negative predictions were further examined to identify systematic failure modes and inform future improvements (Table 8). These failures were not randomly distributed but clustered into two clinically coherent categories, providing structured insight into the conditions under which CTG-FRAME (and CTG-based AI approaches more broadly) remains challenged. The dominant failure mode (60%) involved severe signal loss, characterized by prolonged dropouts, fragmented traces, or extended periods of uninterpretable FHR. In such cases, the remaining visible signal segments frequently appeared reassuring, while pathological events may have occurred during periods of data absence. This pattern reflects an intrinsic limitation of CTG as a monitoring modality rather than a modeling error: in the absence of clinically meaningful signal, neither human experts nor automated systems can reliably infer fetal status. This finding motivates the development of explicit signal-quality modeling and uncertainty-aware escalation strategies as essential extensions for real-world deployment. The remaining false negatives (40%) involved atypical pathological patterns, including recurrent severe decelerations, fetal tachycardia, and chaotic or sinusoidal waveforms. These patterns are clinically recognized but rare and systematically under-represented in publicly available datasets such as CTU-UHB. Their concentration among false-negative predictions illustrates a well-established challenge in rare-event learning: even high-performing models exhibit limited generalization to pattern classes encountered infrequently during training. This limitation is addressable through targeted data enrichment, physiologically informed augmentation strategies, and prospective collection of rare but high-risk CTG phenotypes. Critically, the structure of these failure modes reinforces confidence in the internal validity of the model’s decision logic. The observed errors are attributable to known data and distributional constraints rather than to spurious correlations or unstable decision boundaries. In aggregate, Table 8 complements the quantitative results by demonstrating that CTG-FRAME fails in clinically interpretable and mechanistically explicable ways, strengthening its credibility as a decision-support system rather than an opaque classifier.

Table 8.

Qualitative analysis of false negative cases (N = 10)

Failure Category	Description	Count	Percentage
Atypical Pathological Patterns	Tracings displayed rare but clinically critical patterns, such as recurrent severe decelerations, fetal tachycardia, or chaotic or sinusoidal waveforms. The model likely failed due to a lack of sufficient exposure to these infrequent patterns in the training set, leading to poor sensitivity.	4	40%
Severe Signal Loss	Tracings were characterized by extensive periods of signal dropout, prolonged gaps, or extreme fragmentation. The model failed because the remaining readable segments appeared benign, and there was insufficient continuous data to identify pathology that may have occurred during the gaps.	6	60%

Several research priorities emerge from these findings. First, while internal and external calibration assessments provide encouraging evidence of robustness, prospective multi-center validation remains an essential next step. Such studies should encompass diverse institutions, patient populations, and clinical practice contexts, and should evaluate CTG-FRAME within operational intrapartum workflows to re-estimate performance characteristics under conditions of routine clinical use. This is particularly important for rare, high-consequence outcomes such as severe acidemia, where retrospective benchmarks alone cannot capture the effects of deployment on escalation behavior, clinical decision-making, or patient outcomes.

Second, the integration of electronic health record (EHR) contextual data represents a clinically meaningful and technically tractable extension. Maternal vital signs, parity, gestational age, concurrent medications and relevant comorbidities all influence the clinical interpretation of CTG patterns in practice. Incorporating such covariates into the model architecture could improve probability calibration, provide context for borderline traces, and enhance the system’s capacity to distinguish true risk of acute intrapartum acidemia from benign physiological variation, particularly in cases where signal-only evidence is ambiguous.

Third, closer integration between CTG-FRAME’s structured multi-task outputs and natural-language clinical explanations represents a promising direction for human-centered deployment. Rather than presenting raw model scores, future implementations could translate outputs into concise, audit-ready clinical summaries aligned with established reasoning frameworks. For example, identifying recurrent late decelerations in the final 30 minutes of a recording alongside associated severity estimates and confidence intervals. Recent advances in large language models offer a technically viable pathway for generating such explanations in a controlled and grounded manner, bridging the inferential gap between algorithmic output and bedside clinical communication [53, 54].

Ultimately, the definitive benchmark for clinical utility is not retrospective discriminative accuracy but whether deployment of CTG-FRAME demonstrably improves clinical decision-making and neonatal outcomes, reducing missed acidemia without increasing rates of unnecessary intervention. Establishing this will require rigorously designed prospective studies, including potentially randomized or stepped-wedge trial designs, to evaluate how clinicians interact with the system and how its outputs influence escalation timing, clinical workload, and patient safety within real-world intrapartum care settings.

Conclusion

CTG-FRAME is an end-to-end deep learning framework developed to analyze intrapartum CTG recordings and facilitate the identification of cases at increased risk of severe fetal acidemia. By integrating risk prediction, severity estimation, and temporal signal attribution, the system generates a structured representation of CTG data that aligns with clinical interpretation of fetal heart rate patterns.

A central finding of this study is that effective model learning in this context requires explicit management of extreme class imbalance. The difficulty-aware, multi-stage curriculum supported stable training and enabled the system to distinguish clinically relevant differences between normal and pathological fetal physiology. In contrast, conventional single-stage training resulted in majority-class prediction. These results demonstrate that, for CTG data, the training strategy is fundamental to achieving clinically meaningful model performance.

Ablation analysis showed that key components of the framework, including self-supervised pre-training and cross-signal integration, contribute to performance across tasks and enable robust analysis of noisy or incomplete physiological signals. The system delivers a multi-level clinical assessment by integrating risk identification, acidemia severity estimation, and localization of signal segments contributing to each prediction. These outputs facilitate timely clinical review and ensure transparency in decision-making.

More broadly, this work demonstrates that advancement in intrapartum clinical AI requires not only model accuracy but also alignment between data characteristics, training strategy, and intended clinical application. CTG-FRAME is intended as a decision-support tool, not an autonomous system. A prospective multi-center evaluation and integration of harmonized clinical context are necessary to assess its impact on clinical decision-making, intervention patterns, and neonatal outcomes, and to ensure safe implementation in routine care.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

AI

Artificial intelligence

AUROC

Area under the receiver operating characteristic curve

Bal Acc

Balanced accuracy

CCC

Concordance correlation coefficient

CI

Confidence interval

CTG

Cardiotocography

CTG-FRAME

Cardiotocography foundation risk assessment with multimodal encoding (or your defined expansion)

CTU-UHB

Czech Technical University in Prague and the University Hospital in Brno

EHR

Electronic health record

F1

F1 score

FHR

Fetal heart rate

FS

Sampling frequency

GAP

Gap threshold (signal reconstruction parameter)

Hz

Hertz

GradCam

Gradient-weighted class activation mapping

IoU

Intersection over union

mIoU

Mean intersection over union

MoCo

Momentum contrast

OOF

Out-of-fold

pH

Potential of hydrogen (arterial blood pH)

ResNet-TCN

Residual network for spatial feature extraction with a temporal convolutional network

RMSE

Root mean squared error

SD

Standard deviation

SMOTE

Synthetic minority oversampling technique

SPaM

(Your dataset: define exactly as in manuscript)

STARD-AI

Standards for reporting diagnostic accuracy associated with artificial intelligence studies

SQI

Signal quality index

TRIPOD+AI

Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis statement associated with artificial intelligence studies

UC

Uterine contractions

UCI CTG

University of California, Irvine cardiotocography dataset

W0 / W1 / W2

Training stages in curriculum learning

Youden’s J

Youden’s index

Author contributions

DT made substantial contributions to the conception and design of the study, formal analysis, investigation, methodology, and creation of new software used in the work. DT contributed to drafting the original manuscript and substantively revised it during review and editing. EIY made substantial contributions to formal analysis, investigation, software development, validation, and visualization. EIY contributed to drafting the original manuscript and substantively revised it during review and editing. FDSC made substantial contributions to investigation, validation, and interpretation of the results, and provided substantive critical review and revision of the manuscript with respect to clinical accuracy and relevance. All authors reviewed the manuscript.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The datasets supporting the conclusions of this article are available in the PhysioNet repository as the CTU-UHB Intrapartum Cardiotocography Database, https://physionet.org/content/ctu-uhb-ctgdb/1.0.0/, and in the SPaM dataset, available for academic research use at https://doi.org/10.1111/aogs.13639. A web-based demonstration of the CTG-FRAME framework, including trained models and a user interface, is available at https://ctg.responsibleai.au/. Source code and trained model artefacts supporting the conclusions of this article can be made available to researchers upon reasonable request, under appropriate agreements, as the CTG-FRAME framework is currently under provisional patent review.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that the CTG-FRAME framework is currently subject to provisional patent review. This did not influence the study design, data analysis, interpretation of results, or reporting. The authors declare no other competing interests.