Fetal Health Prediction From Cardiotocography Recordings Using Kolmogorov–Arnold Networks

Abstract

Goal: Cardiotocograph (CTG) is a widely used device for monitoring fetal health during the labor phase. However, its interpretation remains challenging due to the complex and nonlinear nature of the data. Therefore, this paper aims to propose a reliable machine learning model for predicting fetal health. Methods: This paper introduces a state-of-the-art approach for predicting fetal health from CTG recordings (statistical features) using the Kolmogorov-Arnold Networks (KANs). KANs have recently been proposed as a powerful competitor to the conventional transfer function approach in feedforward neural networks. The proposed method leverages the powerful capabilities of KANs to model the intricate relationships within the CTG data, leading to improved classification accuracy. We validate our approach on a publicly available CTG dataset, which consists of statistical features of the acquired recordings and labeled fetal health conditions. Results: The results show that KANs outperform traditional machine learning models, achieving average classification accuracy values of 93.6% and 92.6% for two-class and three-class classification tasks, respectively. Conclusion: Our results indicate that the KAN model is particularly effective in handling the nonlinearity inherent in CTG recordings, making it a promising tool for enhancing automated fetal health assessment.