Improving Explainability in Clinical Mortality Prediction Using Stacking Classifiers over Annotated Clinical Notes

Heart failure, a leading cause of hospitalization and death, requires early identification of high-risk patients to improve care. This study focuses on predicting 3-month mortality after hospital discharge using real-world clinical notes. For such predictive systems to be trusted by clinicians, their decisions must be interpretable and grounded in identifiable medical evidence. We investigate how a stacking classifier can leverage information from automatically annotated clinical notes. Each sub-model focuses on a specific clinical entity type, and a meta-classifier aggregates their outputs to form a final decision. Our experiments show that this yields results (ROC-AUC=0.76) competitive with a global text classifier (ROC-AUC=0.78) while improving interpretability.