Abstract

Preeclampsia is a major cause of maternal and perinatal morbidity and mortality worldwide, and early identification of high-risk pregnancies remains a clinical challenge, particularly in settings where advanced biomarkers and imaging techniques are not readily available. This study aimed to develop and evaluate a machine learning model for multiclass classification of preeclampsia risk levels using routinely available maternal, clinical, and fetal parameters. A publicly available dataset including 203 pregnant women was analyzed. Variables encompassed obstetric characteristics, maternal demographics, medical history, clinical measurements, and fetal parameters. An Extreme Gradient Boosting (XGBoost) algorithm was applied within a multiclass classification framework to categorize preeclampsia risk as low, mid, or high. Model performance was evaluated using accuracy, balanced accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1-score. The model achieved an accuracy of 0.932 in the training dataset and 0.829 in the independent test dataset, with F1-scores of 0.926 and 0.784, respectively. Feature importance analysis identified history of hypertension, proteinuria, systolic blood pressure, and hemoglobin as the most influential predictors, whereas maternal age, parity, and fetal weight showed comparatively lower importance. These findings suggest that an XGBoost-based model using routinely available clinical variables can provide an effective approach for preeclampsia risk classification. However, external validation is required before clinical implementation.

Key words: Preeclampsia, artificial intelligence, risk prediction, extreme gradient boosting, maternal health, pregnancy complications