Machine learning-based prediction of preeclampsia: a comparative study on model effectiveness

Technical Summary: Machine Learning-Based Prediction of Preeclampsia in Low-Resource Settings

Problem Statement
Preeclampsia remains a critical cause of maternal and perinatal morbidity, particularly in low- and middle-income countries (LMICs) like Sri Lanka. While machine learning (ML) has shown promise in predicting this condition, existing models often rely on large datasets from high-income settings and specialized biomarkers unavailable in routine LMIC clinical practice. This study addresses a specific methodological gap: how to effectively develop prediction models under severe data constraints typical of early-phase cohort studies in LMICs. These constraints include small sample sizes, extreme class imbalance (preeclampsia is rare), limited predictor variables (routine clinical features only), and the risk of data leakage when using post-outcome variables. The study aims to identify the most robust algorithm for preeclampsia prediction under these realistic, resource-limited conditions to inform the development of a larger, AI-powered population-based birth cohort in Sri Lanka.

Methodology
The study utilized a retrospective dataset of 343 pregnant women from the Western Province of Sri Lanka, comprising 310 preeclampsia-negative and 33 preeclampsia-positive cases (9.6% prevalence). To simulate a genuine early-prediction scenario and prevent target leakage, the model was restricted to six baseline features available at early pregnancy registration: maternal age, body mass index (BMI), gravidity, parity, number of previous pregnancies, and number of previous births.

Six machine learning algorithms were comparatively evaluated:

1. Logistic Regression (LR)
2. Random Forest (RF)
3. XGBoost (XGB)
4. Support Vector Machine (SVM)
5. K-Nearest Neighbours (KNN)
6. Artificial Neural Network (ANN)

The experimental design employed a rigorous five-fold cross-validation strategy with strict leakage prevention. A critical component of the methodology was the handling of class imbalance. The study compared "baseline" models (using raw imbalanced data or class weighting) against models augmented with the Synthetic Minority Over-Sampling Technique (SMOTE). SMOTE was applied exclusively to the training data within each fold to generate synthetic minority samples, ensuring the validation set remained untouched and representative of the original distribution. Performance was assessed using metrics appropriate for imbalanced classification, including balanced accuracy, Area Under the Curve (AUC), recall (sensitivity), precision, F1-score, and stability (measured as standard deviation across folds).

Key Results

• Baseline Performance: Without imbalance handling, all models failed to detect the minority class effectively. Logistic Regression, KNN, and XGBoost achieved zero recall, essentially predicting the majority class for all samples. While some models (e.g., SVM with class weighting) showed modest recall, they suffered from low precision or extreme variance.
• Impact of SMOTE: Applying SMOTE dramatically improved minority-class detection across all algorithms.
• Model Comparison:
  • XGBoost with SMOTE emerged as the optimal model, achieving the best balance of performance and stability. It recorded a balanced accuracy of 0.889 ± 0.027, an AUC of 0.956 ± 0.013, and a recall of 0.88 ± 0.03. Crucially, it demonstrated the lowest variance (standard deviation) across all metrics, indicating high stability.
  • KNN with SMOTE achieved comparable point estimates (AUC 0.914) but exhibited significantly higher variance (recall SD 0.10 vs. 0.03 for XGBoost), suggesting overfitting given the small sample size.
  • Random Forest showed robust performance (AUC 0.902) but was outperformed by XGBoost in stability and overall metrics.
  • SVM achieved moderate performance (AUC 0.833) with high variance.
  • ANN underperformed relative to tree-based methods (AUC ~0.81), confirming that deep learning offers no advantage for small tabular datasets with limited information content.
  • Logistic Regression performed the worst (AUC 0.733), indicating that the relationship between the selected features and preeclampsia is non-linear.
• Data Leakage Check: Models incorporating post-outcome variables (e.g., gestational age at delivery) produced artificially inflated performance (AUC > 0.98), validating the necessity of strict temporal feature selection.

Key Contributions

1. Algorithm Selection for Data-Constrained Environments: The study provides empirical evidence that ensemble methods, specifically XGBoost combined with SMOTE, are superior to deep learning (ANN) and linear models (LR) for preeclampsia prediction when sample sizes are small (<1,000) and events are rare (<50).
2. Stability over Point Estimates: The paper establishes a methodological framework prioritizing model stability (low variance) over raw point estimates. It demonstrates that high-performing but unstable models (like KNN in this context) are clinically untrustworthy when event counts are low.
3. Feasibility Benchmarking: It sets a realistic performance benchmark for preeclampsia prediction using only routine clinical variables in an LMIC setting, distinguishing this work from studies relying on specialized biomarkers.
4. Rigorous Validation Protocol: The study highlights the critical importance of strict cross-validation protocols that prevent data leakage from synthetic samples (SMOTE) into validation sets, a common pitfall in medical ML research.

Significance and Claims
The authors explicitly position this work as a preparatory methodological feasibility study rather than a clinically deployable tool. The significance lies in its role as a prototype for a planned, larger population-based birth cohort in Sri Lanka.

The paper claims that while XGBoost with SMOTE shows promising discrimination (AUC ~0.96), the current results are not sufficient for clinical use due to:

• The small sample size (only 33 events), which limits statistical power and generalizability.
• The lack of external validation.
• The absence of calibration assessment.

The authors conclude that for the planned birth cohort, future work must focus on accumulating adequate event numbers (targeting 5,000–10,000 participants), conducting external validation, incorporating longitudinal predictors, and performing calibration analysis before any clinical implementation can be considered. The study serves to guide algorithm selection and stability-focused evaluation for future iterations of the AI-powered cohort.