An Attention-Enhanced 3D-CNN Framework for Spectrogram-Based EEG Analysis in Epilepsy Detection

Epilepsy is a widespread neurological disorder affecting approximately 50 million individuals globally, with a disproportionately high burden in low- and middle-income countries. It is characterized by recurrent seizures caused by sudden and uncontrolled electrical discharges in brain cells, often leading to cognitive and motor impairments. Electroencephalography (EEG) remains the gold standard for epilepsy diagnosis, offering non-invasive monitoring of brain activity. However, the complexity and variability of epileptic patterns make traditional visual analysis subjective, time-consuming, and impractical for continuous monitoring. This underscores the need for automated deep learning solutions to improve diagnostic accuracy and operational efficiency. This study introduces a novel 3D Convolutional Neural Network with integrated Self-Attention (3D-ACNN), specifically designed for the volumetric analysis of multi-channel EEG data. The proposed algorithm preprocesses 21-channel EEG signals by segmenting them into 4-second Hann windows, applies the Short-Time Fourier Transform (STFT) to extract time-frequency features, and constructs stacked spectrograms that serve as three-dimensional inputs to the model. The proposed model (3D-ACNN), integrated with self-attention mechanisms, effectively captures spatiotemporal dependencies, enhancing seizure classification accuracy. The model achieved a classification accuracy of 99.89.