Multi-Seizure Classification Using Transformer- Based Attention Models and Pseudo-Labeling Techniques in EEG Analysis

Traditional classification models of seizures from electroencephalogram (EEG) data often struggle to achieve high accuracy in multi-seizure classification due to their complex, and non-linear nature. This study proposes a novel approach for multi-seizure classification leveraging transformer-based attention models combined with pseudo-labelling techniques. The proposed architecture integrates a Transformer model with self-attention mechanisms to capture temporal dependencies and intricate patterns within EEG sequences. The research utilizes the Temple University EEG Corpus, which contains over 30,000 EEG recordings from approximately 315 subjects. Among them clinically relevant subset representing four major seizure types is extracted and analyzed. Experimental results indicate that the model achieved an overall accuracy of 99.2% on the test dataset. Comparative analysis with traditional models demonstrated a significant improvement of up to 10% in classification accuracy and a reduction in false positives by 15%. Cross-validation further verified the stability of the model, with an average cross-validated accuracy of 98.7% and a standard deviation of 0.5%, confirming minimal overfitting. This research introduces an advanced deep learning pipeline capable of transforming the scenario of seizure detection in EEG analysis by offering state-of-the-art accuracy and reliability. Future work may explore real-time deployment and expansion to more diverse EEG datasets to further validate the model’s efficacy.