A Robust and Interpretable Ensemble Learning Framework for Early Mortality Risk Stratification in Heart Failure

Abstract

Heart failure remains a formidable global health challenge, frequently complicated by cardiorenal syndrome, which necessitates early and dynamic mortality risk stratification. Existing clinical scoring systems fail to capture complex nonlinear biomarker interactions, whereas state of the art deep learning models suffer from high computational overhead, algorithmic opacity, and a propensity for severe overfitting on small scale tabular data. To address these critical gaps, this study proposes a computationally efficient and transparent ensemble machine learning framework utilizing Extreme Gradient Boosting (XGBoost) coupled with Shapley Additive Explanations (SHAP). The methodology was rigorously benchmarked against a comprehensive suite of ten distinct machine learning architectures using the publicly validated UCI Heart Failure Clinical Records dataset. Comprehensive evaluations demonstrate that the XGBoost framework is both statistically robust and computationally superior. An ablation study confirmed that synthetic resampling critically maximized diagnostic recall across all evaluated models to prevent fatal misclassifications. Calibrated to an optimal clinical decision threshold (τ = 0.35), the model effectively balanced sensitivity with specificity. Furthermore, five fold cross validation confirmed its supreme generalization stability, achieving a peak mean Area Under the Curve of 0.907 with the lowest algorithmic variance (±0.027) among the ten evaluated models, successfully highlighting the vulnerability of unregularized neural networks and classical algorithms on restricted medical datasets. Game theoretic SHAP analysis biologically validated the framework by decisively isolating elevated serum creatinine and reduced ejection fraction as the primary drivers of mortality. By amalgamating sub four millisecond inference latency with mathematically rigorous clinical interpretability, the proposed framework provides a robust, real time screening tool to optimize proactive interventions in acute cardiovascular care.