A Machine Learning Framework for Preeclampsia Prediction at Isidro Ayora Hospital, Ecuador

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Public Health

Background:

• Preeclampsia (PE) is a significant global cause of maternal and perinatal complications.
• Fragmented clinical data in Ecuadorian public hospitals hinders predictive modeling for PE.
• Limited demographic data availability poses challenges for developing accurate PE prediction models.

Purpose of the Study:

• To develop and evaluate a custom automated Machine Learning (C-AutoML) framework for predicting preeclampsia (PE).
• To assess the performance of various ML models using demographic data from a public hospital in Ecuador.
• To compare the C-AutoML framework against a weighted ensemble model and analyze feature importance.

Main Methods:

• A custom automated Machine Learning (C-AutoML) workflow was developed, evaluating six ML models.
• Performance was compared against a weighted ensemble model using Amazon SageMaker Canvas, optimizing for F1-score.
• Shapley Additive Explanations (SHAP) were used for model interpretability, and a synthetic dataset was generated for analysis.

Main Results:

• The SageMaker ensemble achieved an F1-score of 0.805 and AUC of 0.956 on hospital data.
• The C-AutoML best single model (LightGBM) showed comparable performance (F1-score: 0.802, AUC: 0.944).
• Patient age was identified as the most influential predictor; Logistic Regression achieved perfect classification on synthetic data.

Conclusions:

• The C-AutoML framework provides a competitive and interpretable alternative to ensemble models for PE prediction.
• The methodology demonstrates potential for supporting PE diagnosis in resource-limited settings with demographic data.
• Future integration of clinical variables is expected to enhance the predictive accuracy and risk stratification capabilities for PE.