Hybrid CNN transformer framework for EEG-based epileptic seizure detection

Early diagnosis and early intervention are important in the treatment of epilepsy, so detecting epileptic seizures from EEG signal is very important. However, the non-stationary nature of EEG signals, inter-subject variability and artefacts arising from noise are major challenges to the development of a robust, generalisable and real-time automated seizure detection system. Many current deep learning methods use handcrafted features or rely on single-stream architectures, which fail to adequately account for both spatial and temporal aspects of EEG. Moreover, only a few model feature optimisation techniques compromise model scalability across heterogeneous clinical datasets. EEGSeizureSense is a hybrid deep learning framework that combines deep spatiotemporal feature extraction and hybrid feature optimisation with a CNN-Transformer-based framework for accurate and reliable seizure detection. The statistical, time-frequency, and neighbourhood-informative features of EEG are first extracted, and then a hybrid statistical feature selection technique based on ReliefF and MI is applied. These optimised representations are then fed to the layers for spatio-temporal dependency modelling: the convolutional layers are used for spatial feature learning, while the attention layers are Transformer encoder layers with multi-head self-attention for temporal dependency modelling and contextual representation learning. On the CHB-MIT and TUH benchmark EEG datasets, the experimental evaluation shows an average accuracy of 96.85%, precision of 96.90%, recall of 96.77%, F1-score of 96.81%, and the AUC score of 98.21%. The proposed approach is further verified for robustness and consistency by ablation studies and statistical significance analyses. EEGSeizureSense offers a scalable, clinically applicable and interpretable framework for real-time seizure monitoring and intelligent clinical decision support systems.