Feature Selection-Based Machine Learning Framework for Birth Weight Prediction: Toward Intelligent Clinical Decision Systems

Birth weight is a vital indicator of pregnancy progress and infant development. Accurate pre-delivery estimation is crucial, influencing both short- and long-term health outcomes. This study advances intelligent healthcare systems by proposing a hybrid stacked ensemble model for enhanced birth weight classification. Using a prospective dataset of 913 medical records—each with 18 features including maternal data and fetal ultrasound measurements—we address the challenge of reliable prediction in complex clinical data. Our computational approach incorporates 1) class imbalance correction via Synthetic Minority Oversampling Technique (SMOTE), applied exclusively to training data; 2) mutual-information-based feature selection to identify key predictors; and 3) a hierarchical hybrid architecture uniquely combining support vector machines (SVM), k-nearest neighbors (k-NN), and bagged trees, with a logistic regression meta-classifier. Using 5-fold cross-validation, the hybrid model achieved a weighted-average accuracy of 99.29%, sensitivity of 99.27%, and F1 score of 99.24%, outperforming various standalone models—especially in predicting low birth weight (LBW) and high birth weight (HBW) classes. Results confirm the hybrid model’s value in developing robust, intelligent clinical decision support systems for birth weight classification.