Predicting Maternal Health Risks Using Nutritional Data and Machine Learning

Abstract

Introduction: Maternal health remains a critical global issue, with high mortality rates in low-resource areas due to delayed risk detection and limited healthcare access. Despite medical progress, preventable conditions like hypertensive disorders and gestational diabetes persist, highlighting the need for early diagnostic tools aligned with SDG 3. This study develops machine learning models using clinical data (e.g., blood pressure, glucose) to predict maternal risks, aiming to (1) identify key predictors, (2) evaluate model performance, and (3) support clinical decisions. Challenges include data privacy and quality. The methodology emphasizes preprocessing, model training (XGBoost, KNN), and interpretability for practical deployment, advancing AI-driven solutions for maternal care and SDG 3. Objectives: This study develops machine learning models to predict maternal health risks using clinical indicators like blood pressure and glucose levels. It compares XGBoost, KNN and Random Forest algorithms, evaluating their performance through accuracy, precision and recall metrics. The research identifies key predictive features while examining how data preprocessing affects results. The goal is to create an interpretable risk prediction tool that balances accuracy with clinical usability, particularly for low-resource settings. Implementation addresses data privacy compliance and EHR integration to support healthcare decision-making and improve maternal outcomes. Methods: The study utilized the Maternal Health Risk Dataset, comprising 1,014 entries with features like age, blood pressure, and blood sugar levels. Data preprocessing included outlier removal, encoding, and scaling. Three models—XGBoost, K-Nearest Neighbors (KNN), and Random Forest—were trained and evaluated using accuracy, precision, recall, and F1-score. Hyperparameter tuning was performed via GridSearchCV. Results: The Random Forest model outperformed others, achieving 86.70% accuracy with standardized full features. It excelled in identifying high-risk cases (96% precision, 95% recall). XGBoost followed closely (86.21% accuracy), while KNN lagged (80.30%). Partial feature sets reduced performance across all models. Conclusions: The Random Forest model is recommended for deployment due to its high accuracy and interpretability. Future work includes expanding datasets and integrating real-time EHR systems to enhance predictive capabilities and maternal healthcare outcomes.