A Foundation Model Approach for Fetal Stress Prediction During Labor From cardiotocography (CTG) recordings

TL;DR

This study tackles the variability and limited predictive performance of intrapartum CTG by adopting a foundation-model paradigm with self-supervised masked pre-training on 2,444 hours of unlabeled CTG data using a PatchTST transformer with channel-asymmetric masking. After pre-training, the model is fine-tuned on the CTU-UHB labeled subset to classify fetal acidemia (pH $< 7.15$) during labor, aided by a sliding-window inference scheme for real-time alert generation. The approach achieves an AUC of $0.83$ on the full test set and $0.853$ on uncomplicated vaginal deliveries, outperforming prior benchmarks (${0.68-0.75}$) and indicating that the model captures meaningful CTG–contraction dynamics; error analysis shows most false positives align with clinically relevant patterns despite normal $pH$. The work demonstrates data-efficient learning for CTG and provides reproducible benchmark splits and weights, highlighting the practical potential of foundation models to support labor-room decision-making while acknowledging the need for broader, multi-center datasets and integration of clinical context. $7.15$ is used as the acidemia threshold, and the reported AUC/accuracy numbers reflect clinically relevant performance improvements.

Abstract

Intrapartum cardiotocography (CTG) is widely used for fetal monitoring during labor, yet its interpretation suffers from high inter-observer variability and limited predictive accuracy. Deep learning approaches have been constrained by the scarcity of CTG recordings with clinical outcome labels. We present the first application of self-supervised pre-training to intrapartum CTG analysis, leveraging 2,444 hours of unlabeled recordings for masked pre-training followed by fine-tuning on the 552-recording CTU-UHB benchmark. Using a PatchTST transformer architecture with a channel-asymmetric masking scheme designed for fetal heart rate reconstruction, we achieve an area under the receiver operating characteristic curve of 0.83 on the full test set and 0.853 on uncomplicated vaginal deliveries, exceeding previously reported results on this benchmark (0.68-0.75). Error analysis reveals that false-positive alerts typically correspond to CTG patterns judged concerning on retrospective clinical review, suggesting clinically meaningful predictions even when umbilical pH is normal. We release standardized dataset splits and model weights to enable reproducible benchmarking. Our results demonstrate that self-supervised pre-training can address data scarcity in fetal monitoring, offering a path toward reliable decision support in the labor room.

Figures (6)

• Figure 1: Cardiotocography (CTG) recording examples. Top: Clinical CTG display as seen by clinicians on bedside monitors, showing fetal heart rate (FHR, red, upper panel) and uterine contractions (UC, black, lower panel) on standard grid paper (1 cm = 1 min). Visual interpretation requires simultaneously tracking baseline, variability, accelerations, decelerations, and their temporal relationship to contractions. Bottom: Annotated CTG recording from the FHRMA dataset boudet2019fhrma, illustrating expert consensus annotations of the FHR baseline (purple), accelerations (green shading), and decelerations (pink shading). Even with explicit annotations, classification of these features shows substantial inter-observer variability in clinical practice.
• Figure 2: Examples of UC signal quality issues in CTG recordings. Green shaded regions indicate detected no-value segments where the signal appears flat at non-zero levels due to sensor displacement. Purple shaded regions indicate original NaN values. These flat segments, while numerically valid, contain no physiological information and were excluded from analysis.
• Figure 3: PatchTST architecture overview, adapted from patchtst2023. (a) Channel-independent processing: each channel (FHR, UC) is processed separately through the transformer backbone and concatenated for prediction. (b) Supervised learning: input series is normalized, segmented into patches, projected with positional embeddings, processed by the transformer encoder, and mapped to output via a linear head. (c) Self-supervised pre-training: random patches are masked and the model learns to reconstruct them, enabling representation learning from unlabeled data.
• Figure 4: Illustration of the channel-asymmetric masking strategy. Top: original CTG signal with FHR (blue) and UC (red) channels. Middle: masked input where selected FHR patches are zeroed (gray regions) while UC remains intact. Bottom: binary mask pattern indicating masked (gray) and unmasked (green) FHR patches. The first and last patches are never masked, and masked patches appear in contiguous groups.
• Figure 5: Sliding window inference and alert generation. Example CTG recording with acidemic outcome (pH = 7.02). Top: Model prediction over time, with the 0.5 threshold indicated by the dashed red line. Light gray shading marks regions where predictions exceed the threshold, representing potential alerts. Cyan rectangles illustrate the 7.5-minute sliding windows used to compute each prediction point. Middle: Fetal heart rate (FHR) signal with background color bands indicating normal (green/yellow) and abnormal (pink) ranges. Bottom: Uterine contractions (UC). The shaded region on the right indicates the second stage of labor.