Interpretable prediction of neonatal mortality and its key predictors using machine learning and SHAP analysis
Open Access
- 21 May 2026
- journal article
- Published by Springer Nature in BMC Medical Informatics and Decision Making
- Vol. 26 (1)
Abstract
Introduction: Neonatal mortality is still a global public health concern, especially in lower- and middle-income countries like Ethiopia. Machine learning (ML) has shown excellent performance in healthcare prediction tasks due to its capability of efficiently handling interactions within mixed-type tabular data. Therefore, this study aimed to develop an interpretable model that can predict neonatal death based on the Ethiopian Demographic Health Survey (EDHS) dataset. Methods: This study utilized the imbalanced EDHS dataset collected from all surveys conducted in 2000, 2005, 2011, 2016, and 2019. We evaluated some basic and tree-based ensemble ML algorithms validated with five-fold cross-validation and 80/20 data splitting. Unbalanced, weighted, and SMOTENC class balancing techniques are utilized across both evaluation strategies. The models were compared based on sensitivity and SHAP interpretability with consideration of F1-score and AUC-PR trade-offs. Results: The weighted LightGBM model achieved the highest recall performance (recall = 87.2%) with a slight reduction of F1 = 85.3% and PR-AUC = 92.6% by preserving SHAP interpretability. The SHAP analyses indicate that breastfeeding initiation, number of living children, and ANC visits are the top three key predictors. The dependence plot shows positive SHAP values for delayed breastfeeding initiation, no living children, no antenatal care, lower household member size, grand multiparity, male sex, small birth size, twin births, and short preceding birth intervals. Conclusion: The weighted LightGBM model not only had better sensitivity with competitive AUC-PR results but also showed reliable interpretability using SHAP analyses. The model predicted strong neonatal mortality odds associated with delayed breastfeeding initiation, absence of living children, and lack of antenatal care visits. The model also predicted that neonatal mortality would be more likely for low household member size, grand multiparity, male sex, small birth size, twin births, and short preceding birth intervals. Finally, adequate ANC visits and early initiation of breastfeeding were protective and affected survival outcomes for neonates, emphasizing the importance of maternal and newborn healthcare programs.Keywords
This publication has 76 references indexed in Scilit:
- Optimizing machine learning models for classification of stroke patients with epileptiform EEG pattern: the impact of dataset balancing techniquesProcedia Computer Science, 2024
- A comparison of imputation methods for categorical dataInformatics in Medicine Unlocked, 2023
- Solving the multicollinearity problem to improve the stability of machine learning algorithms applied to a fully annotated breast cancer datasetInformatics in Medicine Unlocked, 2022
- Regularized target encoding outperforms traditional methods in supervised machine learning with high cardinality featuresComputational Statistics, 2022
- Breastfeeding in a Global Context: Epidemiology, Impact, and Future DirectionsClinical Therapeutics, 2022
- Birth order, parental health investment, and health in childhoodJournal of Health Economics, 2021
- Impact of breastfeeding on mortality in sub-Saharan Africa: a systematic review, meta-analysis, and cost-evaluationEuropean Journal of Pediatrics, 2020
- Benchmark for filter methods for feature selection in high-dimensional classification dataComputational Statistics & Data Analysis, 2020
- Effects of preceding birth intervals on neonatal, infant and under‐five years mortality and nutritional status in developing countries: evidence from the demographic and health surveysInternational Journal of Gynecology & Obstetrics, 2005
- Institute of Mathematical Statistics2005