A multidimensional ensemble pipeline for early detection of IUGR condition through CTG

Introduction: Intrauterine growth restriction (IUGR) is a major cause of perinatal morbidity and mortality, often associated with placental insufficiency and progressive alterations in fetal autonomic regulation. Cardiotocography (CTG) represents one of the most widely used tools for fetal monitoring, yet its interpretation remains challenging due to high inter-observer variability and the subtle nature of early pathological patterns. Artificial intelligence approaches have recently shown promising potential for automated CTG analysis, but their development is often limited by the scarcity of large, annotated datasets. Methods: In this study we propose a multidimensional ensemble pipeline for the detection of IUGR from antepartum CTG recordings. The framework integrates two complementary predictive branches: a residual deep learning model (ResNet) operating directly on multivariate temporal sequences, and a hybrid CNN–MLP architecture combining image-based encodings of fetal heart rate signals with physiologically interpretable quantitative descriptors. The outputs of the two models are fused through a logistic regression meta-classifier using a stacking strategy. The pipeline was trained and evaluated using the NAPAMI database, a large clinically curated dataset comprising more than 70,000 CTG recordings collected over a period of 17 years. Results: Both base models (ResNet and CNN + MLP) achieved comparable performance levels. The proposed ensemble approach significantly improved the overall performance, reaching a balanced accuracy of 0.799 and an AUC of 0.868 (95% CI: 0.849–0.885). Statistical comparison using McNemar's test confirmed that the ensemble classifier significantly outperformed the individual models ( p < 10 −11 ). Discussion: The results demonstrate that combining complementary representations of fetal heart rate dynamics through an ensemble framework can improve the detection of IUGR from antepartum CTG recordings. The use of a large-scale clinical dataset together with physiologically informed and deep learning-based representations provides a promising direction for the development of AI-assisted decision support tools in prenatal medicine.