Technical Summary: PRISM-CTG

Problem Statement

Automated Cardiotocography (CTG) analysis using supervised deep learning is currently constrained by the scarcity of high-quality, labeled datasets. Existing models are typically trained on narrowly curated subsets of recordings aligned with specific clinical outcomes (e.g., neonatal acidosis), leading to the exclusion of vast amounts of physiologically informative data that do not fit strict label categories. This reliance on small, task-specific cohorts results in models that lack robustness and fail to learn generalizable, domain-level representations. Furthermore, conventional training pipelines often discard valuable patient metadata and domain-engineered features, treating them as noise rather than potential supervisory signals. While Self-Supervised Learning (SSL) has shown promise in other physiological domains, it has not been effectively applied to CTG, and existing SSL approaches often overlook the rich clinical context inherent in these recordings.

Methodology

The authors propose PRISM-CTG (Physiology-aware Representation Learning via Integrated Self-supervision and Metadata for CTG), a foundation model designed to learn transferable domain-level representations from large-scale unlabelled CTG data. The model is pretrained on over 250,000 hours of CTG recordings (51,336 pregnancies) spanning both antepartum and intrapartum periods.

Core Architecture

PRISM-CTG utilizes a Transformer-based architecture enhanced with three specific design choices to integrate signal and clinical context:

1. Multi-View Pretext Objectives: The model jointly optimizes three complementary objectives:

   • Best-RQ-MAE (Random-Projected Guided Masked Signal Reconstruction): A masked autoencoder approach where a frozen random-projection tokeniser assigns discrete pseudo-labels to signal patches. These labels act as "morphological hints" to guide the reconstruction of masked patches, providing signal-level structural guidance without requiring a deep decoder.
   • Patient Metadata Prediction: A regression task predicting continuous patient variables (gestational age, time to birth, maternal age) to enforce the learning of recording-level, clinically contextual representations.
   • Engineered Feature Classification: A classification task predicting pseudo-labels derived from 17 handcrafted clinical features (e.g., baseline FHR, short-term variability) to align learned representations with established domain knowledge.

2. Task-Specific Tokens and Controlled Cross-Attention:

   • Instead of a single [CLS] token, the model prepends three dedicated learnable tokens: CLS_r (reconstruction), CLS_v (metadata), and CLS_f (features).
   • Isolation Attention Mask: During encoding, each CLS token attends only to itself and the patch tokens, forcing independent aggregation of information specific to its objective.
   • Cross-Attention: After encoding, a cross-attention mechanism allows controlled information exchange between the three CLS tokens, mimicking the clinical process where signal characteristics, patient context, and features inform one another.

3. Signal Tokenization: The input CTG (Fetal Heart Rate and Uterine Activity) is processed via a CNN backbone for patch embedding and a parallel frozen random-projection quantizer for generating discrete label embeddings. These are fused with positional encodings to form the input tokens.

Downstream Evaluation

The pretrained encoder is evaluated using linear probing (freezing the encoder and training a linear head) across seven downstream tasks: four antepartum (abnormality, C-section, NICU, and near-labour classification) and three intrapartum tasks (acidosis classification on CTU-UHB and APHP datasets, and C-section classification).

Key Results

PRISM-CTG was evaluated against a comprehensive set of baselines, including state-of-the-art supervised models (e.g., NeuroFetalNet, PatchCTG) and various SSL foundation models (e.g., MAE, DINO, JEPA).

• Overall Performance: PRISM-CTG achieved the highest average AUC of 0.808 across all seven tasks, consistently outperforming all baselines. The strongest SSL baseline (MAE) achieved 0.773, and the best supervised model (NeuroFetalNet) achieved 0.771.
• Data Efficiency: In low-data regimes (10% of training data), PRISM-CTG outperformed the best supervised baseline (ResNet) by 16.41% and the best full-data supervised model (NeuroFetalNet) by 20.18%. It achieved comparable performance to models trained on full datasets using only 75% of the labeled data.
• Generalization: The model demonstrated strong cross-institutional generalization. On external validation tasks (CTU-UHB and APHP-CTG), it outperformed the best supervised baselines by 2.76% and 8.62%, respectively. Notably, this was achieved despite intrapartum data constituting less than 1% of the pretraining corpus.
• Robustness: PRISM-CTG maintained stable performance under increasing levels of signal dropout (missing data), whereas supervised models showed significant degradation when dropout exceeded 25%.
• Ablation Studies: Removing the Best-RQ-MAE objective caused the largest performance drop (7.16%), highlighting its importance for robust representation learning. Removing the clinical multiview objectives reduced performance by 3.95%, confirming the value of integrating metadata and features.

Significance and Claims

The authors claim that PRISM-CTG represents the first large-scale foundation model for CTG capable of learning transferable domain-level representations. The study's primary significance lies in demonstrating that:

1. Large-scale SSL is viable for CTG: It effectively leverages unlabelled data to overcome the limitations of scarce labeled datasets.
2. Clinical Context is a Supervisory Signal: By reframing patient metadata and domain-engineered features as prediction targets rather than discarded inputs, the model learns clinically grounded representations that are more robust and transferable than those learned from morphology alone.
3. Superior Generalization: The model achieves performance comparable to studies trained on substantially larger, privately labeled datasets, despite using a smaller, publicly available pretraining set and minimal intrapartum data.

The paper concludes that future automated CTG analysis should shift from narrow, task-specific supervised optimization toward large-scale, clinically informed self-supervised pretraining as the primary strategy.