PRISM-CTG: A Foundation Model for Cardiotocography Analysis with Multi-View SSL

PRISM-CTG addresses a key limitation in automated CTG analysis: supervised deep learning models are constrained by small, narrowly curated labelled datasets. To overcome this, the model employs a multi-view self-supervised framework that jointly optimizes three complementary pretext objectives: random-projected guided masked signal reconstruction, clinical variable prediction, and feature classification. By reframing underutilized patient metadata and domain knowledge as prediction targets, PRISM-CTG transforms readily available clinical information into additional supervisory signals, enabling clinically meaningful representation learning from large volumes of unlabelled recordings.

Extensive experiments across 7 downstream CTG tasks in both antepartum and intrapartum domains show that PRISM-CTG consistently outperforms in-domain and SSL baselines. Notably, it demonstrates strong generalization under external validation on 2 separate datasets and achieves performance comparable to studies trained on substantially larger, privately labelled datasets. This is the first study to introduce a large-scale foundation model for CTG that learns domain-level representations, potentially enabling more accurate and scalable fetal health monitoring without the need for expensive labelled data.