The Role of CNN-RNN Hybrid Models and Attention Mechanisms in EEG Signal Recognition for Correct Seizure Detection

Abstract

Epileptic seizure detection is crucial for effective management and treatment of epilepsy. This research proposes a novel hybrid model combining Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and attention mechanisms to enhance the accuracy and reliability of seizure detection from EEG signals. Utilizing the Bonn dataset, our method encompasses advanced preprocessing techniques, including noise reduction and wavelet transforms, to capture multi-scale features from raw EEG data. CNNs extract spatial features, while Bidirectional Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) capture temporal dynamics, with attention mechanisms further refining feature relevance.
To ensure interpretability and trust, Explainable AI (XAI) techniques such as saliency maps, Grad-CAM, and attention maps are integrated. The hybrid model demonstrates superior performance, achieving 95.2% accuracy, 94.1% sensitivity, and 96.5% specificity, significantly outperforming existing methods.
The research highlights the model's robustness through comprehensive evaluation metrics and comparative analysis. Future directions involve testing with diverse datasets, exploring more XAI methods, and real-time implementation. This study advances seizure detection by improving accuracy, interpretability, and clinical applicability, paving the way for enhanced patient care in epilepsy management.