Heart attacks remain the leading cause of mortality globally, highlighting the critical importance of early detection of heart diseases. In emergency diagnostic situations, there is a need for more accurate and sensitive prediction systems that provide quick responses and are accessible anytime, anywhere, even in areas without internet connectivity. However, the existing lightweight machine learning often suffer from reduced accuracy and recall, making them unsuitable for critical applications like heart attack prediction. The dataset is carefully selected for this study to predict severity levels and potential future anomalies of the cardiac disease. To improve the model performance, Synthetic Minority Over-sampling Technique (SMOTE) is considered to address class imbalance. Additionally, a hybrid Feature selection framework is proposed that integrates the Fast Correlation-Based Filter (FCBF) method and the Ant Colony Optimization (ACO) algorithm to eliminate redundant features, followed by hyperparameter tuning through Genetic Algorithm (GA) optimization. The performance of three tree-based algorithms-XGBoost, Random Forest, and Extra Tree Classifier is evaluated with our optimized approach, achieving a maximum accuracy of 98.55%. The proposed work emphasizes that high recall is essential in medical systems, which can be achieved through our novel hybrid feature selection technique. The model is deployed in edge and cloud environments, with edge deployment resulting in a very low response time of 0.0013 s.