MBAT:A Multi-Band Adaptive Transformer Model for Seizure Detection and Seizure Onset Zone Localization from Scalp EEG

We propose a Multi-Band Adaptive Transformer (MBAT) model designed to enhance the accuracy of seizure detection and seizure onset zone (SOZ) localization using electroencephalography (EEG) data. Our model integrates multiband feature extraction with an adaptive transformer architecture to integrate cross-channel information. In order to improve the sensitivity of the model to different EEG channels, we design squeeze-excitation attention (SEAtt) layer to dynamically recalibrates the importance of channels. Then we utilize a gated recurrent unit (GRU) layer for seizure detection. The detection results are used as attention weights in an attention-weighted multi-instance pooling mechanism for SOZ localization. We validate our model through cross-validation on a dataset of 90 patients curated from the Temple University Seizure (TUSZ) corpus. Extensive experiments demonstrate that our model performs better in both seizure detection and SOZ localization tasks. Ablation studies validate the effectiveness of each component, highlighting the significance of our design choices.