<p>Cardiovascular disease (CVD) remains a leading global cause of death, highlighting the need for timely and accurate prediction methods. Traditional diagnostic approaches are often time-consuming, costly, and prone to human error, motivating the use of machine learning (ML) for improved risk assessment. However, ML-based CVD prediction faces challenges such as outliers, irrelevant features, class imbalance, and limited interpretability. In this study, we propose a comprehensive framework that addresses these challenges through advanced data preprocessing, feature selection, data balancing, ensemble learning, and explainable AI (XAI). Outliers are detected and corrected using the Z-score method, while XGBoost-based feature selection identifies the top 9 out of 13 predictors, enhancing model efficiency. Class imbalance is mitigated using techniques such as SMOTE, SMOTETomek, and SMOTE+RUS. Eight ML models - including Logistic Regression, KNN, SVM, Random Forest, XGBoost, Gradient Boosting, AdaBoost, and Gaussian Naive Bayes - are developed and rigorously evaluated using Accuracy, Precision, Recall, F1 score, and AUC-ROC metrics. A voting ensemble combining the top three models (Random Forest, Logistic Regression, and AdaBoost) achieves 98.63% accuracy and 99.13% AUC-ROC on the Cleveland Heart Disease dataset, demonstrating superior performance over individual models. SHAP and LIME analyses provide interpretability, allowing healthcare practitioners to understand model predictions and support clinical decision-making. This work contributes a robust, interpretable, and high-performing ML framework for early CVD detection, offering improved patient risk stratification and potential clinical integration.</p>