PREDICTION OF LENGTH OF STAY AMONG PREECLAMPTIC PATIENTS USING SUPERVISED LEARNING METHODS

Abstract

Hypertensive disorders during pregnancy, particularly preeclampsia, are among the leading causes of maternal and neonatal mortality. In the United States, preeclampsia affects approximately 2 to 8% of pregnancies, with a higher incidence among African American women (6.04%) compared to Caucasian women (3.75%). Due to its severity, preeclampsia often requires intensive care unit (ICU) intervention, resulting in prolonged hospital stays. This study aims to predict the length of stay (LOS) for preeclamptic patients using supervised machine learning on a highly imbalanced dataset. We adopted two modeling approaches: classification and regression, and evaluated multiple algorithms, including logistic regression, decision tree, SVM, KNN, random forest, XGBoost, linear regression, and elastic net. To address class imbalance, we employed oversampling techniques (SMOTE, ADASYN, SMOGN) and cost sensitive learning strategies. Our findings show that cost sensitive logistic regression achieved the highest classification performance with AUC of 66% and G-mean of 60%. Additionally, the analysis revealed that African American women tend to have longer hospital stays. This research supports improved hospital resource allocation, staff planning, and early intervention for high risk cases, contributing to more efficient and equitable healthcare delivery.