An attention-enhanced hybrid CNN-BiLSTM framework for multiclass epileptic seizure recognition using EEG signals.

Introduction: Electroencephalogram (EEG) signal classification plays a pivotal role in brain-computer interfaces (BCIs) and the diagnosis of neurological disorders. However, the non-stationary, non-linear, and low signal-to-noise ratio characteristics of EEG signals make accurate feature extraction challenging. Conventional deep learning approaches often struggle to simultaneously capture localized transient patterns, deep temporal dependencies, and long-range contextual relationships within EEG signals. Methods: To address these challenges, an optimized end-to-end hybrid deep learning framework, termed CNN-BiLSTM-MHSA, is proposed for multiclass epileptic seizure recognition. The architecture sequentially integrates a multilayer one-dimensional Convolutional Neural Network (CNN) for localized feature extraction, a Bidirectional Long Short-Term Memory (BiLSTM) network for capturing forward and backward temporal dependencies, and a Multi-Head Self-Attention (MHSA) mechanism for dynamically emphasizing discriminative neurological events. Robust preprocessing includes Butterworth bandpass filtering, Z-score normalization, and data augmentation using Gaussian noise injection and amplitude scaling. The framework was evaluated on the UCI/Kaggle Epileptic Seizure Recognition dataset comprising 11,500 EEG segments distributed across five classes. Results: The proposed framework achieved an overall accuracy of 98.83%, precision of 98.95%, recall (sensitivity) of 98.84%, F1-score of 98.90%, specificity of 99.71%, balanced accuracy of 98.84%, ROC-AUC of 99.80%, and PR-AUC of 99.72%. Class-wise accuracies exceeded 99% for most categories, with one seizure class achieving perfect recognition. Comparative evaluation further demonstrated that the proposed framework outperformed several recent state-of-the-art EEG seizure classification methods. Discussion: The superior performance can be attributed to the complementary representation capabilities of the hybrid architecture. CNN layers effectively learn localized signal patterns, BiLSTM captures bidirectional temporal dependencies, and MHSA enhances the representation of long-range contextual relationships by assigning greater importance to informative temporal regions. The combination of robust preprocessing and data augmentation further contributes to improved generalization and classification reliability.