Cardiovascular disease represents an ongoing and significant global health concern, and early, accurate cardiac arrhythmia identification is critical to the prevention of severe cardiovascular complications, underscoring the requirement for reliable prediction systems. In this paper, a deep learning-based system for cardiovascular disease classification using the benchmark MIT-BIH Arrhythmia Dataset is proposed, comparing the performance of Gated Recurrent Units (GRUs) with Long Short-Term Memory (LSTM) and one-dimensional Convolutional Neural Network (1D CNN). The utilized dataset is first balanced by class using resampling and separated into subsets (training and testing), and then the training set is normalized (scaled) without leakage to be fed to the deep learning models. To improve generalization and efficiency in training recurrent models, diverse adaptive callbacks are used, including Early Stopping, Reduce LROn Plateau, and Model Checkpoint. The results depicted that although the optimized LSTM and multi-layer 1D CNN models provided strong prediction ability in learning temporal sequence and capturing spatial features of electrocardiogram (ECG) signals, respectively, the optimized GRU model performed superiorly, achieving above 99% accuracy, precision, recall, and F1-score. These results confirm that the GRU model, when incorporated with carefully structured training methodologies, provides an effective and accurate system for predicting cardiovascular disease, with significant possibilities for consolidation into clinical decision support systems.