A multi-stage stacking framework for accurate and interpretable life expectancy modeling

Area of Science:

• Global Health
• Machine Learning
• Health Policy

Background:

• Accurate life expectancy estimation is crucial for global health policy and resource allocation.
• Existing machine learning models lack interpretability and uncertainty quantification.

Purpose of the Study:

• To develop NEXUS-LE, a unified ensemble framework for life expectancy prediction.
• To integrate predictive accuracy, calibrated uncertainty, and policy-level explainability.

Main Methods:

• Trained on global data (22,050 country-year observations, 193 countries, 2000-2021, 150 WHO indicators).
• Employed leave-one-out encoding, PCA, K-Means manifold features, SHAP-guided polynomial expansion, and a three-level stacking architecture.
• Utilized LightGBM residual correction and split-conformal calibration for accuracy and uncertainty.

Main Results:

• Achieved high accuracy on a test set: R²=0.9878, RMSE=1.0665 years, MAE=0.6185 years.
• Conformal prediction provided 95.04% coverage at the 95% nominal level with a 3.32-year interval width.
• Interpretability analysis revealed demographic and structural factors drive 96.3% of attributions.

Conclusions:

• NEXUS-LE offers a reliable and interpretable solution for life expectancy prediction.
• The framework supports evidence-based health policy and global resource allocation.
• Demonstrates the potential of neuro-symbolic AI in advancing global health monitoring.