Predicting Cardiovascular Disease from a Unified Dataset Using Hybrid Data Balancing, Ensemble Learning, and SHAP Explainability

Category	Detail
Condition	Cardiovascular disease (CVD) prediction
Key Mechanisms	Machine-learning classifiers and voting ensembles use a pooled dataset, hybrid class balancing with Instance Hardness Threshold and Tomek Links, and SHAP feature explanations.
Target Population	People represented in the pooled Kaggle, UCI, and Cleveland heart disease datasets; the article does not specify a clinical eligibility population.
Care Setting	Clinical prediction research using publicly available heart disease datasets; the study does not report deployment in a clinical care setting.

The study combines three publicly available heart disease datasets into one pooled dataset.
It applies Instance Hardness Threshold and Tomek Links together to address class imbalance and noisy class boundaries.
Scaling, balancing, and hyperparameter tuning are performed within training folds; held-out test data are not preprocessed.
Twelve classifiers are compared, with soft- and hard-voting ensembles formed from the best-performing classifiers.
Evaluation includes calibration, decision curve analysis, sensitivity–specificity threshold analysis, SHAP explanations, and feature ablation.

The article evaluates machine-learning prediction of CVD using pooled public datasets; it does not provide clinical diagnostic criteria or recommend replacing conventional diagnostic approaches.

The study assesses model performance using calibration and sensitivity–specificity threshold analysis in addition to accuracy.

The authors identify record-level overlap among the merged datasets and absence of external validation as limitations.
The article notes that leakage from partitioning, scaling, resampling, or hyperparameter optimization can make performance appear better than generalization to new patients.

No prescribing population is specified. The analysis uses records from the Kaggle, UCI, and Cleveland heart disease datasets.

The study evaluates prediction models and does not report medication, dosing, or other treatment data.

Perform scaling, resampling, and hyperparameter tuning within training folds, without preprocessing held-out test data.
Use nested cross-validation for hyperparameter optimization, as described in the study.
Assess model performance beyond raw accuracy, including calibration and decision curve analysis.
Report the dataset overlap and lack of external validation as limitations.

Clinical Scorecard: Predicting Cardiovascular Disease from a Unified Dataset Using Hybrid Data Balancing, Ensemble Learning, and SHAP Explainability