Hybridized artificial intelligence system for reducing neonatal mortality in Nigeria

Odeyemi, Charity S.; Olaniyan, Olatayo M.; Omodunbi, Bolaji A.; Samuel, Ibitoye B.; Soladoye, Afeez A.; Olawade, David

Hybridized artificial intelligence system for reducing neonatal mortality in Nigeria

Odeyemi, Charity S. ORCID: https://orcid.org/0000-0002-1160-9394, Olaniyan, Olatayo M. ORCID: https://orcid.org/0000-0002-7349-7500, Omodunbi, Bolaji A., Samuel, Ibitoye B., Soladoye, Afeez A. and Olawade, David ORCID: https://orcid.org/0000-0003-0188-9836 (2026) Hybridized artificial intelligence system for reducing neonatal mortality in Nigeria. International Journal of Medical Informatics, 206. p. 106162.

[thumbnail of 1-s2.0-S138650562500379X-main.pdf]

Preview

Text
1-s2.0-S138650562500379X-main.pdf - Published Version
Available under License Creative Commons Attribution.
| Preview

Official URL: https://doi.org/10.1016/j.ijmedinf.2025.106162

Abstract

Background
Neonatal diseases represent the leading cause of death in Nigeria, ranking the country second globally in neonatal mortality rates. Early and accurate diagnosis remains challenging, leading to delayed interventions and increased mortality.
Aim
To develop an artificial intelligence system capable of detecting multiple neonatal diseases using local datasets and advanced machine learning techniques to facilitate early intervention and reduce neonatal mortality in Southwest Nigeria.
Methods
Clinical records from 4,027 previously treated neonatal patients were collected from five tertiary hospitals across three Southwest Nigerian states. The dataset underwent comprehensive analysis, balancing using Synthetic Minority Over-sampling Technique (SMOTE), and preprocessing before training three deep learning models: Artificial Neural Networks (ANN), Long Short-Term Memory (LSTM), and a novel hybrid LSTM-ANN architecture. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics with rigorous subject-wise validation and statistical testing.
Results
The hybrid LSTM-ANN model demonstrated superior performance with 82 % accuracy, 88 % precision, 82 % recall, and 86 % F1-score, significantly outperforming both standalone ANN (80 % accuracy) and LSTM (77 % accuracy). Disease-specific classification revealed exceptional performance for sepsis (precision: 0.90, F1-score: 0.88), birth asphyxia (0.88, 0.85), jaundice (0.86, 0.83), and prematurity (0.82, 0.80). McNemar’s test confirmed significant hybrid superiority over ANN (χ2 = 12.45, p < 0.001) and LSTM (χ2 = 15.67, p < 0.001), whilst Friedman test (χ2 = 18.42, p < 0.001) validated the 5–6 % accuracy improvement.
Conclusion
The hybrid LSTM-ANN model establishes a valuable diagnostic tool for early neonatal disease detection. However, external validation and prospective clinical trials are necessary before clinical deployment.