Uncovering complex correlations between multidimensional factors and breastfeeding duration using XGBoost

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Reproductive Health

Background:

• Breastfeeding duration is influenced by numerous interacting factors.
• Traditional models struggle with complex, non-linear relationships in clinical data.
• Deep learning models risk overfitting on small datasets.

Purpose of the Study:

• To develop an interpretable machine learning framework for predicting breastfeeding duration.
• To address limitations of traditional and deep learning models in small clinical datasets.
• To identify synergistic risk factors influencing lactation support.

Main Methods:

• Utilized eXtreme Gradient Boosting (XGBoost) on data from 210 postpartum women.
• Employed Explicit Polynomial Feature Engineering for high-order interactions.
• Applied SMOTE for class imbalance and SHAP for feature attribution.

Main Results:

• Achieved 76.2% accuracy and 0.72 AUC, with 0.92 recall for long-term breastfeeding.
• Identified non-linear risk patterns, including delivery mode and education interactions.
• XGBoost outperformed traditional models in predicting breastfeeding duration.

Conclusions:

• An interpretable XGBoost framework effectively models small-sample clinical data.
• Uncovered hidden synergistic risk factors for precision lactation support.
• Advocates for targeted interventions based on identified high-risk profiles.