Reducing neonatal mortality is very important priority in global health, more in regions where medical resources is limited. This problem is very common in developing countries. In this research, machine learning (ML) methods are used to try predicting fetal health by using cardiotocography (CTG) data. The dataset comes from Kaggle and is very useful because many healthcare places not have full access to important medical data. To make models better, Recursive Feature Elimination (RFE) is used to find most important features, like baseline fetal heart rate, uterine contractions, and variability measures. These features were standardized to keep same consistency and trust for all models. Six ML models were made: Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Gradient Boosting (GB), CatBoost (CB), and XGBoost (XGB). They were tested using cross validation. Model results were checked with Matthews Correlation Coefficient (MCC) and F1-score. Between all models, CatBoost (CB) had best MCC score of 0.6452 with 17 features. But it need more computation, so it is hard to use in low-resource places. On other side, Gradient Boosting (GB) got MCC score of 0.6423 with only eight features, giving good balance between accuracy and efficiency. These findings highlight the importance of careful feature selection to optimize both predictive power and practicality. By focusing on models with fewer features, this study presents a viable approach for fetal health monitoring in resource-limited settings. Ultimately, this method contributes to early identification of high-risk pregnancies, aligning with global efforts to reduce neonatal mortality and improve maternal and child health outcomes.