<p>Deep learning (DL) models have demonstrated strong performance in electrocardiography (ECG) arrhythmia classification. However, the criteria and processes for selecting ECG samples for analysis are often unclear. This lack of transparency, together with the challenge of interpreting the rationale behind model predictions, may limit clinical acceptance. This work proposes Dual Attention-based Spatiotemporal Network (DASTNet-X), an explainable spatio- temporal dual attention network to identify discriminative information in both spatial and temporal dimensions of ECG signals. This network employs a multi-domain optimization strategy, progressively refining temporal block objectives to stabilize representation across time, frequency and attention domains. This proposed model integrates dual multi-head attention; a dot product based fusion mechanism and a multi-context feature learning to prioritize representation of salient ECG characteristics. Explainability is demonstrated via attention heatmaps that highlight physiologically significant regions, such as QRS complex. Qualitative evaluations confirm that temporal-domain objectives yield coherent, stable representations, enhancing feature learning and model transparency. Extensive experiments are conducted using incremental dataset partitions ranging from 10% to 90% to evaluate robustness under varying data availability. The proposed work achieves a maximum classification 97.8% while maintain stable performance across all data scales. Performance improvements over baseline models are statistically validated by non-parametric Wilcoxon signed-rank and Friedman tests. The results indicate that the proposed network attains high classification accuracy and improves model interpretability by employing attention-based feature weighting. This approach offers enhanced insight into the decision-making process and demonstrates strong statistical reliability.</p>