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

Abstract

Accurate life-expectancy estimation is critical for evidence-based health policy, monitoring Sustainable Development Goal 3 (SDG 3), and guiding global resource allocation. Existing machine-learning approaches typically provide point predictions with limited interpretability and uncertainty quantification. To address these limitations, we propose NEXUS-LE (Neuro-Symbolic Explainable Unified Stacking for Life Expectancy), a unified ensemble framework that integrates predictive accuracy, calibrated uncertainty, and policy-level explainability. The model is trained on a global dataset comprising 22,050 country-year observations across 193 countries (2000-2021) and 150 World Health Organization indicators. The framework combines leave-one-out encoding, principal component analysis, and K-Means-based manifold features with SHAP-guided polynomial expansion and a three-level stacking architecture, followed by a LightGBM residual correction and split-conformal calibration. On a held-out test set (n = 3,308), the model achieves R2 = 0.9878, RMSE = 1.0665 years, MAE = 0.6185 years, and MAPE = 0.9239%. Conformal prediction attains 95.04% empirical coverage at the 95% nominal level with a 90% interval width of 3.32 years. Interpretability analysis shows that demographic and structural factors contribute 96.3% of total attribution, with consistent non-linear effects across policy domains. The proposed framework provides a reliable and interpretable solution for life-expectancy prediction in global health applications.