Cardiovascular diseases (CVDs) account for most deaths worldwide and require precise, interpretable, and efficient automated diagnostic tools. Although electrocardiography (ECG) is a fundamental tool for non-invasive heart disease diagnosis, manual diagnosis is a laborious and subjective task, and most deep learning approaches lack interpretability and do not apprehension long-range dependances in ECG signals. In this work, an interpretable electrocardiogram image classification method using a customized Tiny Swin Transformer is proposed and demonstrates the effectiveness of a parameter-efficient transformer in achieving cardiologist-level accuracy without the complexity and computational requirements of vision transformers. A structured preprocessing approach optimizes ECG morphology, and a lightweight hierarchical transformer with shifted window self-attention captures both local waveform and global rhythm patterns in ECG signals. Interpretability is obtained using the LIME method, which provides pixel-level attribution maps to identify clinically relevant regions contributing to the predictions. This method was tested on a public electrocardiogram image dataset, which contained 12-lead ECG images from four classes: MI, PMI, Arrhythmia, and Normal. The proposed method has an accuracy of 98.22%, with fewer parameters compared to other architectures. In addition, it has good precision, recall, and F1 scores. The explanations provided by LIME include physiologically relevant features such as ST segment deviation and rhythm irregularity. The framework has good accuracy, efficiency, and interpretability. The framework is useful for real-time automated cardiac screening.