Abstract
This paper considers the challenge of clinical reliability and generalization in heart disease diagnosis, where traditional ensembles often overlook feature heterogeneity and lack uncertainty quantification. We propose an uncertainty-aware feature-weighted ensemble (UAFE) framework designed for robust cardiovascular prediction. UAFE integrates three core components: (1) Feature Stratification to train specialized learners on importance-based subgroups; (2) Dynamic uncertainty weighting to adaptively adjust model contributions based on prediction disagreement; and (3) Neighborhood refinement to rectify decisions for high-uncertainty samples using local geometric priors. Comprehensive experiments on a multi-center dataset demonstrate that UAFE outperforms state-of-the-art baselines with an accuracy of 0.8660. Furthermore, leave-one-center-out (LOCO) validation across four international clinical sites shows that UAFE maintains superior stability (mean accuracy 0.8332) against institutional distribution shifts, establishing its effectiveness for deployment in unseen clinical environments.