Development and Internal Validation of Preeclampsia Risk Prediction and Stratification Models: Conventional Regression and Machine Learning Approaches in Zambia

Objectives: To examine associations between maternal factors and preeclampsia, and to develop and internally validate any-onset preeclampsia risk prediction and stratification models. Design: Cross-sectional study with prediction model development and internal validation. Setting: Fourteen facilities across four provinces of Zambia. Population: 15,385 pregnancies recorded between 2019 and 2024. Methods: Multivariable logistic regression assessed associations between maternal factors and preeclampsia. Logistic Regression (LR), Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models were developed using an 80/20 train/test split, class weighting and nested three-fold cross validation. Performance was assessed using discrimination, calibration and risk stratification. Outcome measure: Preeclampsia spectrum disorders. Results: Preeclampsia prevalence was 2.5%. Chronic hypertension (aOR 14.09, 95% CI 8.80-22.55; p<0.0001) and history of hypertension in pregnancy (aOR 3.60; 95% CI 2.21-5.88; p<0.001) were strong predictors of preeclampsia. Parity was protective: parity 1-4 (aOR 0.36; 95% CI 0.25-0.51; p<0.00) and >5 (aOR 0.36; 95% CI 0.21-0.62; p<0.001). There was no evidence that maternal age ≥35 years (p=0.858) and malaria (p=0.624) were independently associated. RF achieved the highest AUROC (0.874; 95% CI 0.826-0.921) while LR (AUROC 0.845; 95% CI 0.768; 0.902) showed stable calibration. Risk stratification demonstrated increasing event rates across risk groups (p<0.001). Conclusion: Logistic Regression offers a practical, interpretable approach for preeclampsia risk prediction and stratification in low-resource settings using routine data. External validation is required.