Truss bridges play a critical role in transportation infrastructure, valued for their efficiency in carrying heavy loads across long spans. However, despite being designed with high redundancy, they face various challenges that can compromise their structural integrity and potentially result in local damage and collapse. For this, maintaining their safety, stability, and long-term serviceability requires an advanced structural health monitoring (SHM) process, capable of detecting damage at an early stage. This study investigates AI-assisted damage detection methods in truss bridges, with particular emphasis on single-member failure, through the application of Graph Neural Networks (GNNs). GNNs provide a principled framework for incorporating structural topology directly into data-driven damage identification models. The case study involves a validated finite element (FE) model of a scaled steel truss bridge, used to simulate all possible member failure scenarios under specified loading, and generate corresponding structural response datasets. These datasets include both transitional displacement and modal displacement responses to characterize damage-sensitive features. The combined use of such responses enables the capture of complementary damage-sensitive characteristics, reflecting both localized stiffness degradation and global dynamic behavior resulting from a member failure. In the proposed framework, the truss bridge is represented as a graph, in which nodes correspond to structural joints and edges represent structural members. Translational and modal displacements are assigned as node-level features based on a consistent mapping between measured degrees of freedom and nodal coordinates. Member failure states are defined at the edge level and formulated as a binary multiclass classification problem, in which each member is labeled as either failed or intact. Hence, the GNN will learn spatial correlations between nodal response patterns and member-level failure using iterative message transmission while keeping the underlying structural connectivity, enabling effective learning under sparse measurement conditions. The findings of this study will demonstrate the potential of integrating a graph-based learning framework to enhance robustness, scalability, and interpretability in truss bridge damage detection. The suggested approach will provide a promising pathway toward data-driven SHM systems that integrate physics-based modeling with advanced deep