Hand pose estimation is a critical area of research in multi-modal human-computer interaction within the contemporary information era. While extant algorithmic models demonstrate notable accuracy in estimation, their deployment in real-time interactive systems is often impeded by computational complexity and extensive parametric requirements. To address this limitation, we propose HiSAT-Net, an optimized algorithm for RGB image-based hand pose estimation that balances speed and accuracy. The proposed architecture leverages a lightweight encoder-decoder as its core, employing a hierarchical structure for down-sampling and up-sampling processes to facilitate resolution recovery and keypoint localization. We integrate the Distribution-Aware Coordinate Representation (DARK) methodology to refine keypoint positional information. Additionally, we introduce a novel metric, the Speed-Accuracy Trade-off (SAT) index, to quantitatively assess the model's performance equilibrium between computational efficiency and estimation accuracy. Our approach innovatively combines hierarchical algorithmic structures with the Shuffle Attention (SA) mechanism, while incorporating DARK for keypoint position refinement. Experimental evidence suggests that this synergistic integration achieves superior speed-accuracy performance with a relatively compact parametric footprint. This study contributes to the field by demonstrating the efficacy of a hierarchical model in maintaining high-quality hand pose estimation while significantly reducing computational demands, thereby enhancing its applicability in real-time interactive systems.