A Hybrid Deep Learning Approach for Seizure Classifcation Using CNN-BiLSTM and MASFNet

Abstract

Epilepsy is a neurological disorder that affects many individuals worldwide. Accurate seizure classification using electroencephalogram (EEG) signals is essential for supporting epilepsy diagnosis. However, EEG signals are often noisy, rapidly varying, and characterized by patterns occurring at multiple temporal scales. This work proposes a hybrid deep learning model that integrates CNN-BiLSTM and MASFNet for multiclass seizure classification. The CNN-BiLSTM branch captures temporal dependencies using multi-resolution convolution and bidirectional long short-term memory, while the MASFNet branch extracts multi-scale spectral-spatial features through parallel convolution combined with Squeeze-and-Excitation (SE) attention. The features learned from both branches are merged to produce final classification output using fully connected layers. The proposed model is evaluated on the BEED dataset, which contains 8000 EEG samples distributed across four seizurerelated classes. During training, mixup augmentation, focal loss, and cosine decay learning rate scheduling are employed to improve generalization. A 5-fold cross-validation strategy is used to ensure reliable evaluation. The proposed model achieves 98.19% accuracy, 98.22% precision, and 98.19% F1-score, demonstrating strong performance for automated seizure classification.