Machine learning based classification in obstetrics: evaluating models, partitioning strategies, and key predictors in cardiotocography.

Detecting fetal distress early and accurately is critical to the reduction of adverse neonatal outcomes. However, this proves to be a challenging task. The use of artificial intelligence (AI) provides a promising solution by facilitating the analysis of cardiotocography (CTG) signals in a more timely and precise manner. Despite this, AI in fetal monitoring faces several critical obstacles. In this study, we evaluated six experimental scenarios on a publicly available UCI CTG dataset consisting of 2,160 cases, each containing 21 features and examined how selected machine learning-based models, various data partitioning strategies, and input feature selection influence classification performance and predictor importance. We compare multiple classification models—Classification and Regression Trees (CART), C5, Random Tree, Chi-squared Automatic Interaction Detection (CHAID), K-Nearest Neighbor (KNN), Multilayer Perceptron (MLP), Radial Basis Function Network (RBF), and Support Vector Machines (SVM). According to our results, we were able to predict normal, suspect, and pathological cases with an overall accuracy ranging from 86.14% up to 96.92%. Our experiments shed light on how different dataset partitioning strategies affect the results and interpretability of significant predictors. In addition, we tested a scenario based on clinical parameters routinely acquired by our best practices in order to determine how these parameters may affect our model’s performance. Our findings provide practical recommendations for optimizing both the accuracy and interpretability of machine learning-based models. Furthermore, they emphasize the urgent need for larger, more comprehensive datasets as well as standardized definitions of outcomes. These insights aim to encourage clinical centers to create their own high-quality databases and collaborate across institutions, ultimately advancing the adoption of AI-driven decision-support tools in routine obstetric practice.