An Explainable AI Framework for Neonatal Mortality Risk Prediction in Kenya: Enhancing Clinical Decisions with Machine Learning

Neonatal mortality remains a critical public health challenge in Kenya, with a rate of 21 per 1,000 live births—well above the SDG 3.2 target. While machine learning (ML) offers potential for risk prediction, most models lack transparency and clinical interpretability, limiting their adoption in low-resource settings. This study presents an explainable AI (XAI) framework for predicting neonatal mortality using Kenya Demographic and Health Survey (KDHS) data (N = 2,000), with a focus on model accuracy, fairness, and clinical relevance. Six ML models—Logistic Regression (LR), KNN, SVM, Naïve Bayes, Random Forest, and XG-Boost—were trained and evaluated using in-sample, out-of-sample, and balanced datasets, with performance assessed via AUC, F1-score, sensitivity, specificity, and Cohen’s Kappa. To address class imbalance and enhance generalizability, synthetic oversampling and rigorous cross-validation were applied. Post-balancing, LR achieved optimal performance (AUC = 1.0, κ = 0.98, F1 = 0.987), with SVM (AUC = 0.995) and XG-Boost (AUC = 0.982) also showing higher performance. SHAP and model breakdown analyses identified Apgar scores (at 1st and 5th minutes), birth weight, maternal health, and prenatal visit frequency as key predictors. Fairness assessments across socioeconomic subgroups indicated minimal bias (DIR > 0.8). The integration of XAI enhances transparency, supports clinician trust, and enables equitable decision-making. This framework bridges the gap between predictive accuracy and clinical usability, offering a scalable tool for early intervention. Policy recommendations include embedding this XAI-enhanced model into antenatal care systems to support evidence-based decisions and accelerate progress toward neonatal survival goals in resource-limited settings.