An explainable artificial intelligence–based stacking ensemble for heart failure mortality prediction: A data-leakage-free clinical decision support approach

Abstract

Accurate mortality risk prediction at emergency department admission is critical for triage and patient safety in heart failure. We aimed to develop a data-leakage-free stacking ensemble model predicting mortality from baseline variables, targeting high sensitivity for clinical safety, and utilizing Explainable Artificial Intelligence (XAI). A publicly available dataset of 299 heart failure patients was analyzed retrospectively. To prevent data leakage, the follow-up time variable was excluded. Class imbalance was addressed by applying the Synthetic Minority Oversampling Technique (SMOTE) exclusively to the training set. A stacking ensemble model was constructed using Random Forest, XGBoost, and LightGBM as base learners, with Logistic Regression as the meta-learner. To reflect clinical priorities, the decision threshold was optimized solely on the training set to achieve ≥ 85% sensitivity. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). The stacking model achieved an area under the curve (AUC) of 0.838. On the test set, it reached 89.5% sensitivity and 92.3% negative predictive value (NPV), reliably identifying high-risk patients. SHAP analysis confirmed clinical plausibility, showing predictions were primarily driven by acute indicators (serum creatinine, ejection fraction, and age) rather than chronic comorbidities. This data-leakage-free stacking ensemble provides a transparent and reliable decision support tool for heart failure triage. Its high sensitivity and integration of XAI methods can significantly enhance clinician confidence in AI-assisted evaluations.