Cross-subject seizure detection with vision transformer and unsupervised domain adaptation

Abstract

Automatic seizure detection is of critical importance for clinical epilepsy treatment. Due to the variability of Electroencephalography (EEG) patterns across different individuals, most existing seizure detection methods fails to generalize across patients. To tackle this issue, this paper proposes a cross-subject seizure detection combines Vision Transformer (ViT) and unsupervised domain adaptation (UDA). Specifically, to enhance the generalization ability of the ViT backbone across different subjects, an adversarial network is introduced on the class token to disentangle global transferable features. Meanwhile, the multi-head attention mechanism is replaced by the transfer adaptation module (TAM) to disentangle transferable features at the patch level. Additionally, to retain the discriminative features related to epileptic seizures, a discriminative clustering module (DCM) is introduced to constrain the model. Our experiments on the CHB-MIT dataset demonstrate that the proposed method achieves strong performance in both evaluation paradigms: in epoch-based analysis it attains 89.20% accuracy, 91.05% sensitivity, and 94.54% specificity, while in event-based evaluation it maintains 89.23% sensitivity with a low false detection rate of 0.42/h. The results verify the feasibility of this method in cross-subject seizure detection.