Breastfeeding Prediction Using Machine Learning: Insights into Key Predictors and Model Performance.

Abstract

Background: Breastfeeding shows persistent disparities shaped by complex maternal, infant, and household factors, making prediction challenging. As traditional models struggle to capture such complexity, this study aims to identify predictors of ever breastfeeding and longer breastfeeding duration using multiple machine learning models and to assess the potential of these methods in breastfeeding prediction.

Methods: Using nationally representative data from the U.K. Household Longitudinal Study, we apply Lasso (Least Absolute Shrinkage and Selection Operator) logistic regression, decision tree, random forest, XGBoost, support vector machine, and neural network models to examine factors associated with whether infants were ever breastfed and whether breastfeeding stopped at or after 3 months.

Results: Maternal, household, and birth-related factors all relate to breastfeeding, but maternal socioeconomic factors, especially education and ethnicity, emerge as the strongest predictors. Across the two breastfeeding outcomes, model performance is better for whether infants were ever breastfed than for breastfeeding duration, which is more difficult to predict. Within each outcome, the models show similar performance without substantial differences.

Conclusions: Breastfeeding outcomes can be reasonably predicted, and the predictive signal is driven primarily by maternal socioeconomic characteristics, enabling targeted support at relatively low informational cost. However, prediction remains limited, especially for breastfeeding duration, suggesting that while key influences are captured, more nuanced or situational mechanisms remain difficult to anticipate. Lasso logistic regression, with its low complexity and transparent structure, performs comparably to more complex models and may therefore be a more appropriate methodological choice for breastfeeding prediction.