<![CDATA[Electrochemical Impedance Spectroscopy (EIS) has emerged as a promising modality for non-invasive biomedical diagnostics, particularly for radiation-free monitoring tasks such as Bone Mineral Density (BMD) assessment. However, the high dimensionality, noise, and non-linear behavior of impedance signals pose significant challenges for accurate and interpretable prediction. This study introduces Hybrid Rough Set-Attention Network (HRSA-Net), a hybrid regression framework that combines Rough Set-based feature selection with a self-attention neural architecture to enable continuous BMD estimation directly from raw EIS data. The proposed framework employs Artificial Neural Network (ANN) and Transformer-based regression models to learn complex impedance-density relationships. Unlike prior studies that are limited to classification tasks or rely on indirect physiological indicators, HRSA-Net is explicitly designed for direct regression of real-valued BMD scores. The model performance is evaluated against reference measurements obtained from Dual-energy X-ray Absorptiometry (DXA), the current clinical gold standard for bone density assessment. Through a comprehensive series of ablation experiments, HRSA-Net achieves an R² of 0.834 using an attention-guided ANN backbone, demonstrating the critical contribution of both Rough Set reduction and attention mechanisms. Performance further improves to an R² of 0.855 when incorporating a Transformer regressor and Huber loss, indicating superior robustness and generalizability under varying signal conditions. Comparative analysis with state-of-the-art EIS-based learning approaches shows that the proposed pipeline consistently outperforms conventional neural models and statistical methods. Overall, HRSA-Net provides an interpretable, accurate, and scalable foundation for future portable EIS-based BMD diagnostic systems, offering a safer alternative to radiological methods such as DXA and enabling feasible deployment in primary or community healthcare settings]]>
