Epilepsy is a prevalent neurological disorder that poses significant health risks. Electroencephalogram (EEG)-based seizure detection is a widely used technique for epilepsy diagnosis. The existing Graph Neural Network (GNN)-based methods have not considered the semantics and dynamic characteristics. In this paper, a dynamic GNN model with brain region semantics for EEG-based seizure classification is proposed. Specifically, temporal features are extracted to capture dynamic neural patterns, while semantic representations of brain regions are incorporated to encode anatomical information. A unified multi-relational graph is constructed by integrating spatial, temporal, and semantic similarities, subsequently processed by a graph-based classifier for seizure classification. Experimental results on the popular dataset demonstrate that the proposed method achieves higher performance than the compared methods across both binary and multi-class seizure classification tasks.