Meta-cluster driven ensemble learning with cluster-augmented features for explainable heart disease prediction

Heart disease remains a major global cause of death, necessitating accurate and interpretable predictive models. In this study the work proposes a Meta Cluster-Driven Ensemble Learning Engine (MCDELE) to improve predictive accuracy. Our proposed model uses a two-tier stacking architecture. The first layer consists of a range of base learners trained on a combined heart disease dataset collected from Kaggle, which is derived from multiple UCI Heart Disease cohorts. This layer is enhanced with two cluster-augmented features that represent structural patterns within the dataset. Four statistical meta-features are then computed from the base learners’ predictions and passed on to the second meta-learner layer. The proposed ensemble achieved the highest accuracy (90.22%), F1-score (91.26%), and precision (90.38%) while comparing with the baseline models, and attaining a strong AUC of 93.77%. The role of cluster-augmented features and meta-features in the model’s decision is emphasized by SHAP-based interpretation. The statistical analysis using the Wilcoxon signed-rank test supports the competitiveness of MCDELE over several competing models across multiple evaluation metrics. The finding demonstrates that MCDELE is a reliable and interpretable framework for a medical decision support system.