Improved Hybrid Local Binary Structural Pattern Shallow Graph Deep Convolutional Attention Neural Networks With...

Epileptic seizure (ES) detection from electroencephalography (EEG) signals is difficult because of noise and the intricate, patient-specific nature of brain activity. Traditional methods often suffer from low accuracy, high computational costs, and poor generalization. To tackle these challenges, an improved hybrid local binary structural pattern shallow graph deep convolutional attention neural networks with synergistic fibroblast optimization (IHLBSPSGDCAN2Nets + SFO) is proposed for automated ES detection as well as diagnosis in EEG signals. First, the input EEG signals obtained in the Bonn and Children's Hospital Boston and Massachusetts Institute of Technology (CHB-MIT) datasets are subjected to preprocessing using the subaperture keystone transform matched filtering (SAKTMF) approach, which allows reducing noise and artifact effects. Subsequently, feature extraction is performed using the second-order synchroextracting transform combined with empirical wavelet transform (SOSTC-EWT), capturing both time-frequency as well as spectral characteristics of EEG signals with high precision. For classification, the improved hybrid local binary structural pattern shallow graph deep convolutional attention neural network (IHLBSPSGDCAN2Nets) architecture is employed, which integrates the improved local binary pattern shallow deep convolutional neural network (ILBPSDCNN) with a hybrid structural graph attention network (HSGAN). Synergistic fibroblast optimization (SFO) is also employed to optimize hyperparameters and achieve optimal model performance, with accelerated convergence, fewer classification errors, and minimal computational cost. Results from experiments demonstrate that IHLBSPSGDCAN2Nets + SFO delivers an outstanding classification accuracy of 99.9, by far exceeding that of conventional methods. The proposed method effectively handles noise, extracts precise EEG features, optimizes model performance, enhances convergence, and significantly improves classification reliability, robustness, and efficiency, outperforming traditional methods in automated ES detection.