Interpretable prediction of neonatal mortality and its key predictors using machine learning and SHAP analysis

Area of Science:

• Public Health
• Machine Learning
• Predictive Analytics

Background:

• Neonatal mortality remains a significant global public health challenge, particularly in low- and middle-income countries such as Ethiopia.
• Machine learning (ML) offers powerful tools for healthcare prediction by effectively analyzing complex, mixed-type data.

Purpose of the Study:

• To develop an interpretable machine learning model to predict neonatal death.
• Utilize the Ethiopian Demographic Health Survey (EDHS) dataset for model development and validation.

Main Methods:

• Employed various machine learning algorithms (basic and ensemble tree-based) on imbalanced EDHS data (2000-2019).
• Utilized five-fold cross-validation, 80/20 data splitting, and class balancing techniques (unbalanced, weighted, SMOTENC).
• Evaluated models based on sensitivity, F1-score, AUC-PR, and SHAP interpretability.

Main Results:

• The weighted LightGBM model demonstrated superior recall (87.2%) with competitive F1 (85.3%) and AUC-PR (92.6%), maintaining SHAP interpretability.
• Key predictors identified include breastfeeding initiation, number of living children, and antenatal care (ANC) visits.
• Factors increasing neonatal mortality risk include delayed breastfeeding, no living children, no ANC, male sex, and short preceding birth intervals.

Conclusions:

• The weighted LightGBM model provides accurate and interpretable predictions for neonatal mortality.
• Delayed breastfeeding, absence of living children, and lack of ANC visits are strongly associated with increased neonatal death risk.
• Promoting adequate ANC visits and early breastfeeding initiation is crucial for improving neonatal survival outcomes.