Abstract
<title>Abstract</title> <p> <bold>Background:</bold> Preeclampsia remains a leading cause of maternal and perinatal morbidity in low- and middle-income countries (LMICs). As preparatory methodological work for an AI-powered population-based birth cohort in Sri Lanka, this study systematically evaluated machine learning algorithms for preeclampsia prediction under realistic constraints of small sample size, severe class imbalance, and routine clinical features. <bold>Methods:</bold> Data from 343 pregnant women in the Western Province of Sri Lanka were analysed, including 310 preeclampsia-negative and 33 preeclampsia-positive observations. Six algorithms were compared: Logistic Regression, Random Forest, XGBoost, Support Vector Machine (SVM), K-Nearest Neighbours (KNN), and Artificial Neural Network (ANN). Multiple imbalance-handling strategies, including SMOTE, class weighting, and cost-sensitive learning, were evaluated using five-fold cross-validation with rigorous leakage prevention. <bold>Results:</bold> Without imbalance handling, baseline models achieved high nominal accuracy (0.90-0.91) but zero sensitivity for minority-class detection. XGBoost with SMOTE showed the most favourable performance in this exploratory analysis (balanced accuracy: 0.889 ± 0.027; AUC: 0.956 ± 0.013), although estimates remained unstable because of the small number of events (n=33). KNN with SMOTE achieved comparable results (AUC: 0.914 ± 0.045), while Random Forest with SMOTE showed robust performance with lower variance (AUC: 0.902 ± 0.013). SVM with SMOTE achieved moderate performance (AUC: 0.833 ± 0.042). Logistic regression without balancing failed completely (recall=0.0), whereas logistic regression with SMOTE achieved an AUC of 0.733. ANN performance was sample-limited (AUC approximately 0.81). Models incorporating post-outcome variables produced artificially inflated but methodologically invalid performance (AUC >0.98), demonstrating target leakage. <bold>Conclusions:</bold> Ensemble methods with SMOTE can achieve promising discrimination using routine clinical variables, but this work should be interpreted as feasibility methodology rather than a clinically deployable tool. The small sample size, lack of external validation, and absence of calibration assessment preclude current clinical use. For the planned birth cohort, adequate event accumulation, rigorous temporal feature selection, and external validation remain essential prerequisites. </p>