Cardiotocography (CTG) is an established method for monitoring fetal well-being through measurements associated with fetal heart rate and uterine contractions. Automated classification of CTG records into Normal, Suspect, and Pathological categories can support clinicians by providing rapid and consistent decision-support information. Conventional machine-learning techniques, particularly Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), have demonstrated strong performance on publicly available CTG datasets. Deep-learning models such as Long Short-Term Memory (LSTM) networks have also been investigated because of their ability to learn nonlinear representations. This paper proposes a hybrid LSTM–Random Forest framework in which an LSTM network performs representation learning and the learned representation is subsequently classified using Random Forest. The framework is designed for the UCI Cardiotocography dataset, which contains 2,126 CTG records and 21 automatically extracted features with three fetal-state classes. A leakage-resistant experimental protocol is proposed using stratified partitioning, training-only preprocessing, class-imbalance handling, and multiple evaluation metrics including accuracy, balanced accuracy, precision, recall, macro-F1, Matthews correlation coefficient, and ROC-AUC. Explainability is incorporated using SHAP. Published literature provides benchmarks against which the proposed framework can be evaluated. Reported results include RF accuracies of 94.90% and 97.20% under different experimental protocols, XGBoost accuracy of 95.27%, LSTM accuracy of 93.4%, and hybrid/stacked approaches reaching 95.39% and 98.9% accuracy in different studies.

Keywords: Cardiotocography, fetal health classification, fetal heart rate, uterine contractions, Random Forest, Long Short-Term Memory, XGBoost, hybrid machine learning.