<p>The cardiac arrhythmias are diverse ranging from a mild condition to a life-threatening condition and are markers of underlying heart diseases. The ability to detect changes early and accurately before the development of severe cardiovascular event such stroke and heart failure is essential. Despite the fact that electrocardiogram (ECG) continues to be the primary diagnostic tool, traditional analysis is tedious, prone to errors, and hard for complex ECG signal patterns. This research significantly advances automated ECG arrhythmia detection by integrating comprehensive multi-dataset evaluations, employing a Transformer-based deep learning architecture specifically tailored for modeling long-range temporal dependencies, and applying robust interpretability techniques. In addition, we merged several public ECG databases, such as MIT-BIH Arrhythmia Database, PhysioNet challenge database and European ST-T database in order to improve diversity and generalization of our model. The Transformer model based on the self-attention techniques showed better performance over the traditional machine learning and deep learning algorithms (SVM, Random Forest, CNN, and LSTM), reaching 98.3% in accuracy and 0.992 in AUC-ROC. We used Grad-CAM, SHAP and LIME, to provide transparent insights into the model’s diagnostic decisions consistent with clinical knowledge in order to address the interpretability burden by nature of deep learning methods. It also naturally incorporated systematic evaluations of alternative loss functions (Focal Loss, Dice Loss) that significantly improved the performance on minority arrhythmia classes, and these improvements facilitated data imbalanced completely. The innovations presented here collectively deliver an accurate, interpretable, and clinically practical automated system for cardiac arrhythmia detection, representing a clear step beyond existing approaches in terms of performance, robustness, and clinical applicability.</p>