Advanced Persistent Threats (APTs) pose a significant challenge to cybersecurity, as their sophisticated strategies often evade traditional detectors that fail to capture complex temporal and structural patterns in system call sequences. To address this gap, we propose the Graph-Temporal Fusion Network for Intrusion Detection (GT-FID), a novel dual-branch deep learning architecture. GT-FID synergistically integrates a Long Short-Term Memory (LSTM) network to model time-ordered dependencies with a Graph Neural Network (GNN) that analyzes structural relationships within dynamically constructed call graphs. Evaluated on the public ADFA-LD dataset, GT-FID achieves a test accuracy of 0.9622 and a Macro-Averaged F1-Score of 0.95, significantly outperforming strong baselines including GRU (0.9462) and Transformer (0.9563) models. These results demonstrate that fusing temporal and structural features provides a more robust and effective representation for detecting complex attack patterns, establishing a promising direction for future host-based intrusion detection systems.