ESCAViT: Symmetry-Aware EEG Classification

Se Hwan Lim, Hyungyu Lee

Keywords: IIIC Pattern Classification, EEG Transformer, Inter-Lead Contrastive Learning, Class Imbalance Mitigation

TL;DR: We propose ESCAViT, a multi-stream Transformer-based classification framework that addresses inter-lead interactions and class imbalance in EEG time series.

Abstract: Accurate classification of Ictal-Interictal-Injury Continuum (IIIC) patterns is essential for neurological assessment in intensive care units, yet remains challenging due to limitations in capturing inter-lead correlations and addressing class imbalance. To tackle this, we propose ESCAViT, a multi-stream Transformer-based EEG classification framework. ESCAViT leverages the Video Vision Transformer with specialized feature extraction mechanisms to model spatiotemporal EEG patterns, while applying domain-adaptive learning to enhance data diversity and mitigate heterogeneous Other class(HOC) effects. Experimental results on the IIIC dataset show that ESCAViT outperforms state-of-the-art models, achieving 21.9% improvement in mean accuracy per class (mACC) and 22.6% in F1-score. Our method significantly enhances LRDA classification by over over 20%, thereby addressing classification bias. ESCAViT demonstrates consistent performance across different IIIC patterns and imbalanced distributions, confirming its effectiveness in EEG classification. The code is available at https://github.com/limshmai/ESCAViT.git

Submission Number: 7