Machine learning (ML) holds great promise in transforming healthcare by enabling early disease detection and improving clinical decision-making. Despite this potential, ML models often face obstacles such as irrelevant features, overfitting, and limited generalization capabilities. This research addresses these challenges by focusing on cardiovascular disease prediction using a well-known dataset from Kaggle. The study implements a structured optimization approach, beginning with feature selection methods—such as correlation filtering and regularization techniques (Lasso and Ridge)—to reduce redundancy and enhance data quality. To improve model accuracy and control overfitting, we evaluate high-performing ensemble algorithms including Random Forest, XGBoost, and LightGBM. Each model undergoes fine-tuning using hyperparameter optimization and stratified k-fold cross-validation to validate performance across subsets of the data. Our findings confirm that gradient boosting methods deliver superior accuracy and consistency compared to more conventional algorithms. This work emphasizes the value of tailoring ML pipelines to the specific challenges of medical data, demonstrating how thoughtful preprocessing and algorithm selection can significantly enhance prediction outcomes. The study provides a practical framework that can be adapted to other healthcare applications, with the goal of advancing precision medicine in the field of cardiovascular care.