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

PURPOSE

This study aims to evaluate the performance of machine learning models using different data imputation techniques in different balancing scenarios, employing sociodemographic attributes and maternal health history, using data of a population from the state of Pernambuco, Brazil, to predict fetal death during pregnancy.

METHODS

We used a dataset from a social program in Pernambuco, Brazil, covering the period from 2008 to 2022, that includes sociodemographic, prenatal, maternal and family health history data. We separated two scenarios with two balancing techniques to train the models, Random Undersampling (RU scenario) and Hybrid Undersampling 2x (H2X scenario) and we explored using four tree-based machine learning models, each of which was evaluated based on their performance and feature importance.

RESULTS

The models were evaluated under different metrics. The XGBoost model stood out with 81.06% specificity and the Random Forest model stood out with 67.73% sensitivity, in different scenarios. The attributes that most impacted the learning process were first prenatal care, age, education and interpregnancy interval.

CONCLUSION

This application is particularly valuable in the context of social projects, such as those in Brazil, where innovative solutions can contribute to achieving the SDGs offering a unique perspective on the intersection of technology, healthcare, and social impact.