Evaluating how different balancing data techniques impact on prediction of premature birth using machine learning...

Area of Science:

• Medical Informatics
• Machine Learning
• Public Health

Background:

• Premature birth (before 37 weeks gestation) is a leading cause of neonatal mortality globally.
• Predictive modeling for preterm birth faces challenges due to imbalanced datasets, potentially leading to biased outcomes.
• The Brazilian healthcare system (SUS) seeks improved tools for early identification of high-risk pregnancies.

Purpose of the Study:

• To evaluate machine learning models for predicting premature birth using Brazilian data.
• To address data imbalance issues using various sampling techniques.
• To enhance the accuracy of preterm birth prediction for improved clinical intervention.

Main Methods:

• Utilized a dataset of over 483,000 Brazilian sociodemographic and obstetric cases.
• Compared five data balancing techniques: Undersampling, Oversampling, and three Hybrid-sampling configurations.
• Trained and evaluated Decision Tree, Random Forest, and AdaBoost machine learning models.

Main Results:

• Hybrid-sampling techniques significantly outperformed Undersampling and Oversampling in predictive model performance.
• The Decision Tree model with Hybrid-sampling achieved 70% accuracy, 64% recall, and 74% precision.
• Demonstrated the critical role of appropriate data balancing in developing reliable preterm birth prediction models.

Conclusions:

• Hybrid-sampling is a superior approach for balancing imbalanced data in preterm birth prediction models.
• Improved prediction accuracy can facilitate earlier identification of at-risk pregnancies, enabling timely interventions.
• The findings have significant implications for enhancing maternal and neonatal care within the Brazilian Unified Health System (SUS), potentially reducing neonatal mortality.