<p>The early detection of cardiovascular diseases (CVDs) is critical for improving patient outcomes, but traditional electrocardiogram (ECG) datasets suffer from limitations, such as small sample sizes, imbalanced class distributions, and a focus on specific disease types. This paper introduces a novel approach to addressing these challenges by creating a hybrid single-lead ECG dataset derived from the MIT-BIH Arrhythmia, European ST-T, and BIDMC Congestive Heart Failure databases. A dataset was developed through rigorous preprocessing steps—normalization, smoothing, segmentation, and class balancing—to enhance diversity and maintain consistent data quality. To evaluate the dataset’s effectiveness, a five-layer CNN with an attention mechanism was trained, achieving significant improvements across binary and multi-class classification tasks on all evaluation metrics. Additionally, a basic one-layer CNN model was employed for statistical validation, comparing performance on the hybrid dataset versus individual datasets. A paired t-test confirmed statistically significant improvements with low p-values across all metrics. This work establishes a robust foundation for advanced ECG analysis and underscores the potential of balanced, multi-source datasets in overcoming typical limitations of traditional ECG datasets.</p>