Machine learning for preventing stillbirths: is it possible to transform data into life-saving insights?

Area of Science:

• Maternal Health
• Machine Learning
• Public Health

Background:

• Fetal death prediction is crucial for maternal and child health.
• Utilizing sociodemographic and health history data can improve predictive accuracy.

Purpose of the Study:

• Evaluate machine learning model performance for fetal death prediction.
• Compare data imputation techniques and balancing scenarios.
• Identify key predictors of fetal death in a Brazilian population.

Main Methods:

• Utilized a dataset from a social program in Pernambuco, Brazil (2008-2022).
• Employed Random Undersampling (RU) and Hybrid Undersampling 2x (H2X) balancing techniques.
• Evaluated four tree-based machine learning models (e.g., XGBoost, Random Forest) on performance and feature importance.

Main Results:

• XGBoost achieved 81.06% specificity; Random Forest achieved 67.73% sensitivity.
• First prenatal care, maternal age, education, and interpregnancy interval were significant predictors.
• Model performance varied across different balancing scenarios and imputation techniques.

Conclusions:

• Machine learning offers a valuable tool for fetal death prediction in public health initiatives.
• This approach can support social programs in Brazil and contribute to Sustainable Development Goals.
• Integrating technology in healthcare can yield significant social impact.