An EEG-based seizure detection method with multiband guided fusion and cross-frequency interaction

The automated detection of epilepsy is of great importance for improving the efficiency of clinical diagnosis. However, existing methods are insufficient in modeling cross-frequency interactions in multiband electroencephalogram (EEG) signals. To address this limitation, we propose an epilepsy detection model based on multiband guided fusion. The goal is to enhance the modeling capacity of cross-frequency dynamic features through band division and a hierarchical feature integration mechanism. A multi-scale temporal encoding module (MTEM) is designed for each specific frequency band, enabling the extraction of local dynamic features from low-, mid-, and high-frequency signals. Furthermore, a hierarchical fusion framework is developed, which incorporates a cross-frequency guidance mechanism and a frequency-specific statistical attention module (FSAM). This design effectively captures cross-frequency interactions across different frequency bands, which are often overlooked by previous methods that fail to model cross-frequency dependencies. Our model enhances these interactions, outperforming existing methods with an F1-score of 99.42% on the TUSZ dataset, 99.9% on the CHB-MIT dataset, and 95.46% on the TUEV dataset. The results indicate that the strategy of frequency band division and cross-frequency fusion can effectively capture multiscale dynamic patterns related to epilepsy, offering a more reliable automated solution for clinical EEG analysis.