Title: A Lightweight and Interpretable Three-Hidden-Layer Backpropagation Neural Network for Intrapartum Fetal Risk Triage Using Cardiotocographic Features
Authors: Nesreen Samer El_Jerjawi
Samy S. Abu-Naser
Volume: 10
Issue: 7
Pages: 94-104
Publication Date: 2026/07/28
Abstract:
Cardiotocography (CTG) is the standard non-invasive technique for intrapartum fetal monitoring; however, its interpretation remains challenging due to substantial inter-observer variability and the complex interaction among fetal heart rate patterns and uterine activity. This study proposes a lightweight and fully interpretable artificial neural network (ANN) developed entirely within the JustNN environment for automated fetal risk triage using 21 standardized CTG-derived features. The proposed model employs an extremely compact 21?2?1?1?1 feedforward architecture containing only 47 trainable weights. The network was trained using the Backpropagation algorithm with momentum over 46,500 learning cycles on 1,387 training examples and evaluated on an independent validation set of 739 CTG records. The final model achieved a validation accuracy of 90.53%, correctly classifying 669 validation cases while maintaining a negligible computational footprint. Beyond predictive performance, this study provides a comprehensive analysis of the complete learning dynamics and the native feature-importance ranking generated directly by JustNN. The learning trajectory demonstrates stable convergence without observable evidence of significant overfitting, despite extended training beyond the initial convergence stage. Furthermore, the most influential features identified by the model - accelerations, abnormal short-term variability, and uterine contractions - are consistent with established clinical understanding of fetal well-being and distress. Although more complex machine learning approaches such as ensemble models and deep neural networks may achieve higher classification accuracy, they often require greater computational resources and reduced transparency. In contrast, the proposed ANN demonstrates that clinically meaningful fetal risk prediction can be achieved through an extremely compact, reproducible, and interpretable architecture suitable for resource-constrained clinical environments.