Electroencephalography-based epileptic seizure classification using a sparse iterative recursive-gated recurrent unit

Abstract

Epileptic seizure classification using electroencephalography (EEG) plays a significant role in clinical diagnosis, long-term patient monitoring, and treatment planning. However, accurate seizure classification remains challenging due to its high complex and non-stationary nature of brain signal, which leads to misdiagnosis. This research proposes a sparse iterative recursive (SIR)-gated recurrent unit (GRU) to classify epileptic seizures accurately. By incorporating GRU into SIR, the model effectively learns and retains significant temporal patterns in EEG signals over long sequences. The gated structure minimizes the vanishing gradient, resulting in effective learning from sequential seizure patterns. Sparsity and recursion enhance the feature representation with lesser computational complexity for accurate seizure classification. Time-and frequency domain feature extract with sparsity and iterative recursion enables the model to selectively concentrate on seizure relevant temporal patterns, which minimizes redundant and noisy information. Hence, SIR-GRU achieves higher accuracies of 99.75% and 99.54% on the Bonn EEG as well as Children's Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) dataset, respectively, compared to the existing method including spatio-temporal feature fusion dual attention (STFFDA). These findings indicate that SIR-GRU provides rapid and reliable performance, which improves accuracy in clinical practice.