LGFNet: A Local-Global Focused Network for EEG Seizure Detection

Electroencephalography (EEG) serves as a widely adopted non-invasive modality for epilepsy diagnosis and seizure detection. Nevertheless, achieving reliable cross-patient generalization relying solely on algorithms remains a fundamental challenge, due to substantial inter-patient variability, non-stationary background activity, and heterogeneous electrode configurations. To address these issues, we introduce LGFNet, a Local–Global Focused Network designed for robust seizure detection. Central to LGFNet are Local-Focus Adapter which incorporates Re-parameterized (Rep) layer and Layered-Scale Convolution (LSConv), and Global-Focus Attention. In essence, the Rep layer effectively extracts local features, thereby improving the model's adaptability to inter-patient variations in EEG signals. Meanwhile, LSConv addresses the fusion of features using Kernel Applier (KA) and Kernel Predictor (KP) with large-kernel perception and small-kernel aggregation mechanisms, effectively solving the issue of cross-electrode feature integration. Finally, global attention further enhances the generalization ability of the seizure detection. This hierarchical Local Layered-Scale, Connect-Global architecture emphasizes transient localized bursts while ensuring coherent global integration, aligning closely with seizure neurophysiology. We evaluate LGFNet on the widely-adopted CHB-MIT dataset under a rigorous cross-patient protocol, employing 2-second EEG segments sampled at 256 Hz. Experimental results demonstrate that LGFNet attains an accuracy of 97.03%, consistently surpassing competitive baselines while incurring minimal overhead. Comprehensive ablation studies further validate the indispensable roles of both layers in the Local-Focus Adapter, underscoring their synergistic contribution to robust feature extraction and reliable seizure prediction. These results indicate that LGFNet constitutes a promising step toward practical real-time clinical deployment in epilepsy management.