Cardiovascular diseases represent a significant global health challenge, with arrhythmia detection being crucial for early intervention. Although artificial intelligence has been widely applied in ECG-based arrhythmia detection, existing approaches face challenges with imbalanced data and inefficient feature representation. This study introduces ECG-GAT, a novel hybrid feature selection framework that synergistically integrates Graph Attention Networks (GAT) with Sequential Forward Feature Selection and Random Forest-based feature importance ranking to revolutionize electrocardiogram-based arrhythmia detection. Our approach systematically addresses two fundamental limitations plaguing existing systems: the pronounced performance degradation in minority class detection due to dataset imbalance, and the inadequate exploitation of intricate feature relationships through conventional linear representations. By leveraging Principal Component Analysis to strategically decompose the classification problem into two optimized scenarios, the proposed method achieves remarkable improvements in minority class sensitivity while simultaneously reducing computational overhead. Comprehensive evaluation on the MIT-BIH Arrhythmia database demonstrates that ECG-GAT not only outperforms state-of-the-art approaches in accuracy and precision metrics but also maintains exceptional efficiency with reduced feature sets. The ECG-GAT framework offers a promising solution for improving automated arrhythmia detection systems, particularly in handling imbalanced datasets and leveraging complex feature relationships through graph-based representations.