Machine learning prediction of kangaroo mother care in Sierra Leone: a comparative study of feature selection techniques and classification algorithms

Osborne, Augustus; Soladoye, Afeez A.; Usani, Kobloobase O.; Adekoya, Ayomide Israel; Wada, Ojima Z.; Olawade, David

Machine learning prediction of kangaroo mother care in Sierra Leone: a comparative study of feature selection techniques and classification algorithms

Osborne, Augustus, Soladoye, Afeez A., Usani, Kobloobase O. ORCID: https://orcid.org/0009-0009-9668-3128, Adekoya, Ayomide Israel ORCID: https://orcid.org/0009-0006-6275-9030, Wada, Ojima Z. and Olawade, David ORCID: https://orcid.org/0000-0003-0188-9836 (2026) Machine learning prediction of kangaroo mother care in Sierra Leone: a comparative study of feature selection techniques and classification algorithms. International Journal of Medical Informatics, 206. p. 106166.

[thumbnail of 1-s2.0-S1386505625003831-main.pdf]

Preview

Text
1-s2.0-S1386505625003831-main.pdf - Published Version
Available under License Creative Commons Attribution No Derivatives.
| Preview

Official URL: https://doi.org/10.1016/j.ijmedinf.2025.106166

Abstract

Background
Kangaroo Mother Care (KMC) is a critical intervention for improving neonatal outcomes, particularly for low-birth-weight infants. Identifying predictors of KMC practice remains essential for targeted health interventions and policy development.
Objective
This study utilizes data from the 2019 Sierra Leone demographic and health survey to identify predictors of KMC using different feature selection techniques and classification algorithms.
Methods
We analyzed 7,377 maternal and child health records from the 2019 Sierra Leone demographic and health survey, applying three feature selection techniques and seven classification algorithms. Data preprocessing included class balancing and cross-validation. Three feature selection techniques employed were: Adaptive Ant Colony Optimization (ACO), Recursive Feature Elimination (RFE), and Backward Feature Selection. Seven machine learning algorithms implemented were: Logistic Regression, Support Vector Machine variants, K-Nearest Neighbours, Random Forest, XGBoost, Stacking Ensemble, and Voting Ensemble. Data preprocessing included SMOTE for class imbalance, 5-fold and 10-fold cross-validation, and hyperparameter optimization using GridSearchCV.
Results
Random Forest and XGBoost consistently achieved the highest performance across all feature selection methods. Using consensus features from multiple selection techniques, Random Forest achieved an accuracy of 0.72, F1-score of 0.78, and ROC-AUC of 0.7689, whilst XGBoost demonstrated similar performance (accuracy: 0.72, F1-score: 0.78, ROC-AUC: 0.7685). Backward Feature Selection and ACO outperformed RFE in identifying discriminative features. Ensemble methods showed robust generalization capabilities.
Conclusion
Machine learning models, particularly ensemble methods combined with comprehensive feature selection techniques, demonstrate strong predictive capability for KMC practice, offering valuable insights for maternal and child health interventions in Sierra Leone.