Channel Transformer-Based Generative Adversarial Network With Multi-Instance Attention and Nutcracker Optimization...

The current literature on automatic seizure detection based on EEG has obtained significant accuracy, but most of them still have difficulties in processing the highly non-linear, non-stationary, and patient-specific EEG signals. Models typically need vast quantities of training data; they do not generalize to other datasets, and they are sensitive to noise and changes in channels, making them less robust and applicable in clinical practice. To overcome them, this paper will assume a channel transformer-based generative adversarial and multi-instance attention network with a nutcracker optimizer (CTGA-MinsAN-NutO) to identify seizures reliably. The suggested structure incorporates adaptive guided multi-layer side window box filter decomposition (AGM-LSWBFD) to perform well in denoising and multi-directional shearlet transform domain (MDSTD) to carry out more efficient feature extraction. The model outperforms current benchmarks and shows better robustness in identifying ictal and interictal states, achieving 99.1% accuracy and 93.5% recall when evaluated on the Bonn as well as CHB-MIT datasets.