Ensemble Techniques Based Risk Classification for Maternal Health During Pregnancy

Abstract

This research focuses on the critical aspect of maternal health during pregnancy, emphasizing the need for early detection and intervention to address potential risks to both mothers and infants. Leveraging various classification methods, including Naïve Bayes, decision trees, and ensemble learning techniques, the study investigates the prediction of childbirth potential and pregnancy risks. The research begins with data collection, followed by preprocessing to clean and prepare the data, including handling missing values and normalization. Next, cross-validation is performed to ensure model robustness. Five ensemble techniques are used for risk classification: Ensemble Boosted Trees, which enhances the performance of decision trees; Ensemble Bagged Trees, which combines predictions from decision trees trained on different subsets of data; Ensemble Subspace Discriminant, which applies discriminant analysis on random subspaces; Ensemble Subspace KNN, which uses K-Nearest Neighbors (KNN) within random subspaces; and Ensemble RUS Boosted Trees. Key variables such as maternal age, height, Hb levels, blood pressure, and previous pregnancy history are considered in these analyses. Additionally, the study introduces Ensemble Learning based on Classification Trees, revealing significant improvements in accuracy compared to cost-sensitive learning approaches. The comparison of methods, including Naïve Bayes and K-Nearest Neighbor, provides insights into their respective performances, with ensemble techniques demonstrating their potential. The proposed ensemble learning techniques, namely Ensemble Boosted Trees, Ensemble Bagging Trees, Ensemble Subspace Discriminant, Ensemble Subspace KNN, and Ensemble RUS Boosted Trees, are systematically evaluated in classifying pregnancy risks based on a comprehensive dataset of 1014 records. The results showcase Ensemble Bagging Trees as a standout performer, with an accuracy of 85.6%, indicating robust generalization and effectiveness in clinical risk assessment compared to traditional methods such as Decision Tree (61.54% accuracy), K-Nearest Neighbor (74.48%), Ensemble Learning based on Cost-Sensitive Learning (73%), Ensemble Learning based on Classification Tree (76%), Gaussian Naïve Bayes (82.6%), Multinomial Naïve Bayes (84.8%), and Bernoulli Naïve Bayes (84.8%). Ensemble Bagging Trees achieved the highest accuracy proving to be more effective than the other methods. However, the study emphasizes the need for continuous refinement and adaptation of ensemble methods, considering both accuracy and interpretability, for successful deployment in healthcare decision-making. These findings contribute valuable insights into optimizing pregnancy risk classification models, paving the way for improved maternal and infant healthcare outcomes.