Interpretable Machine Learning for Mortality Risk Detection in National Health Data

Background: Accurate mortality prediction is essential for identifying high-risk individuals and guiding public health interventions. However, machine learning (ML) models trained on nationally representative data—such as NHANES, where mortality occurs in fewer than 10% of cases—often struggle with extreme class imbalance and limited interpretability, hindering practical utility. Objective: This study investigates whether loss-aware ML approaches can enhance both sensitivity and interpretability in predicting all-cause mortality, particularly in older and socioeconomically vulnerable populations where most mortality events occur. Methods: We used data from 4,188 U.S. adults in the 2011–2012 NHANES cycle, linked to the 2019 National Death Index. Four models—logistic regression, random forest, gradient boosting, and XGBoost—were trained under varying loss function strategies, without imputation or oversampling, to preserve real-world class imbalance. Performance was assessed via recall, F1-score, and PR-AUC. SHapley Additive ex- Planations (SHAP) were used for interpretability. Mortality label distributions across preprocessing strategies were statistically compared using chi-square tests. Results: Despite modest absolute metrics, the XGBoost model with class-weighted loss achieved the best recall (30.7%) and F1-score (35.4%), enabling identification of a substantial portion of deaths that baseline models missed. SHAP analysis revealed clinically consistent risk factors—age, HbA1c, systolic blood pressure, and poverty index—particularly concentrated in high-risk subgroups. Crucially, chi-square tests showed that both the raw (χ2 = 48.26, p < 0.0001) and SLOTE-imputed datasets (χ2 = 798.81, p < 0.0001) differed significantly from our analytic dataset in outcome distribution, underscoring the importance of rigorous preprocessing. These distortions, if ignored, could bias both model evaluation and real-world risk stratification. Conclusion: Even when predictive accuracy appears modest, transparent and statistically grounded models offer valuable insights for targeted outreach and equitable health policy. Our results demonstrate that interpretable, loss-aware ML can play a critical role in population-level mortality prediction. All code and data will be made publicly available upon peer-reviewed publication.