Objective:

Develop and evaluate an explainable machine-learning approach for cardiovascular disease prediction using a pooled dataset, hybrid class balancing, and leakage-controlled validation.

Approach:

• Data: Pooled three publicly available heart disease datasets: Kaggle, UCI, and Cleveland.
• Balancing and validation: Combined Instance Hardness Threshold and Tomek Links. Scaling, balancing, and nested hyperparameter tuning were performed within training folds; held-out test data were not preprocessed.
• Models and evaluation: Compared 12 classifiers and used soft- and hard-voting ensembles. Evaluation included calibration, decision curve analysis, and sensitivity–specificity threshold analysis, in addition to accuracy.
• Interpretability: Used SHAP to examine feature contributions and conducted a feature ablation study.

Interpretation:

The provided article context describes the study rationale and methods but does not report its performance results.

Limitations:

• The authors identified record-level overlap among the three pooled datasets.
• External validation was not conducted.

Conclusion:

The study proposes a leakage-controlled, explainable ensemble-learning framework for CVD prediction; the supplied context does not include the results needed to assess its predictive performance.

Sources:

• Article context supplied in the prompt