FPMBGCN : Feature pyramid-based multi-branch graph convolutional networks for epileptic seizure detection

Epilepsy is a prevalent neurological disorder, and the automatic detection of seizure periods plays a vital role in clinical diagnosis and treatment. However, significant variability in seizure patterns across patients poses substantial challenges for cross-patient detection, including pattern heterogeneity and noise interference. Traditional convolutional neural networks struggle to model the topological dependencies between EEG channels, while existing graph neural network (GNN) approaches, which typically rely on a single graph input, exhibit limited capacity in capturing inter-channel topological relationships. To address these limitations, this paper proposes the feature pyramid-based multi-branch graph convolutional network (FPMBGCN) for epileptic seizure detection using multi-channel EEG signals. The method constructs three complementary graph-structures: pearson correlation coefficient (PCC), phase-locking value (PLV), and spatial distance (SD) graphs—to extract shared topological features across patients. It further incorporates a channel weight block to adaptively emphasize critical channels, a graph feature pyramid to capture multi-scale information while suppressing noise, and the graph squeeze-and-excitation (GraphSE) module alongside a self-attention mechanism to enhance the representation of spatial and temporal dynamics. The performance of the proposed method is evaluated on both the CHB-MIT dataset and the Siena scalp EEG dataset with leave-one-out cross-validation. For the CHB-MIT dataset, the proposed FPMBGCN achieves an average accuracy of 92.21%,a precision of 94.50%, a recall of 90.03%, an F1 score of 78.59%, and an F1 score of 91.64% in cross-patient detection. On the Siena dataset, it achieves an accuracy of 77.94%, a precision of 76.25%, a recall of 81.37%, a specificity of 74.41%, an F1 score of 78.59%. • FPMBGCN combines multi-branch GCN and feature pyramid for cross-patient EEG seizures. • Constructs three graph structures to represent the topology of EEG signals. • Achieves state-of-the-art classification performance on the CHB-MIT dataset.