patchCTG

Antepartum cardiotocography (CTG) plays a crucial role in monitoring fetal health. The CTG is one of the most common tests performed during pregnancy worldwide. However, traditional methods face limitations, including high inter-observer variability and false positive rates, resulting in poorer outcomes for the pregnancy. We developed PatchCTG, a transformer-based model specifically designed to analyse CTGs. Our approach employs advanced patch-based tokenization, instance normalization, and channel-independent processing, allowing it to effectively capture local and global temporal dependencies in CTG signals.

In our evaluation using the Oxford Maternity dataset, PatchCTG demonstrated strong performance. It achieved an area under the curve (AUC) of 77%, outperforming other methods. PatchCTG balanced sensitivity (57%) and specificity (88%), providing more reliable predictions across diverse clinical conditions. Importantly, the model performed particularly well when fine-tuned with data closer to delivery, showcasing its adaptability to varying temporal thresholds.

The potential of PatchCTG extends beyond its predictive capabilities. By reducing the subjectivity of manual CTG interpretation, it offers a more consistent, objective tool for clinical decision-making. Its ability to adapt sensitivity and specificity thresholds makes it suitable for both high-risk scenarios, where sensitivity is critical, and low-risk contexts, where specificity can help avoid unnecessary interventions. With further validation and integration of additional clinical data, PatchCTG could significantly enhance fetal health monitoring and support better outcomes for mothers and babies.

Read the pre-print here.