Early and accurate diagnosis of cardiac diseases through electro- cardiogram (ECG) analysis has historically been crucial for preventing complications, but it has often been hindered by the challenges of manual interpretation. This paper proposes a novel approach for automated ECG heartbeat classification using multi-step preprocessing and machine learning model optimization, aiming to achieve high accuracy, robustness, and interpretability, potentially surpassing existing techniques. Utilizing the MIT-BIH Arrhythmia Database, which encompasses diverse ECG morphologies, our study involved an extensive preprocessing pipeline including exploratory data analysis, noise reduction techniques (Gaussian smoothing, thresholding), feature extraction (gradient computation, absolute rolling maximum), and dimensionality reduction (decimation, CSR matrix conversion). A comprehensive evaluation of various ma- chine learning models—such as KNN, SVM, decision trees, random forests, AdaBoost, Naive Bayes, logistic regression, one-vs-rest, LSTM, and CNN—was conducted, focusing on rigorous training and evaluation metrics with an emphasis on interpretability and generalizability. Our findings indicate that the optimized convolutional neural network (CNN) model significantly outperformed the other models, achieving a testing accuracy of 98.07% and an F1 score of 0.982, showcasing the effectiveness of our preprocessing and optimization strategies. These results highlight the potential of our approach in enhancing the automated classification of ECG signals, offering a promising avenue for improving the diagnostic process for cardiac diseases.