PatchCTG: A Patch Cardiotocography Transformer for Antepartum Fetal Health Monitoring

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Antepartum Cardiotocography (CTG) is crucial for fetal monitoring but suffers from subjective visual interpretation and low accuracy.
• Existing automated systems like Dawes-Redman have limitations in detecting fetal distress, showing high specificity but low sensitivity.

Purpose of the Study:

• To introduce PatchCTG, an AI-enabled time series transformer designed for objective and accurate Cardiotocography (CTG) analysis.
• To evaluate PatchCTG's performance in identifying fetal distress and enhancing antepartum fetal health assessment.

Main Methods:

• PatchCTG utilizes patch-based tokenisation, instance normalisation, and channel-independent processing for CTG signal analysis.
• The AI model was trained and validated on the Oxford Maternity (OXMAT) dataset, comprising over 20,000 CTG traces.
• Extensive hyperparameter optimisation was performed to enhance model performance.

Main Results:

• PatchCTG achieved an Area Under the Curve (AUC) of 0.77, with 88% specificity and 57% sensitivity at the Youden's index threshold.
• The model demonstrated robust predictive performance across varying temporal thresholds, showing potential for real-time and retrospective analysis.
• Fine-tuning for near-delivery cases yielded 52% sensitivity and 88% specificity.

Conclusions:

• PatchCTG offers a promising AI-driven, sensor-based tool for objective fetal health assessment in antepartum care.
• The system has the potential to improve clinical decision-making by providing reliable fetal well-being evaluations.
• PatchCTG's adaptability and robust performance highlight its significance for advancing fetal monitoring technologies.