Deep Learning–Assisted Hybrid EEG/Non-EEG Architecture for Accurate Seizure Prediction

Epileptic seizures are transient disruptions of cerebral activity that can severely impact quality of life if not detected promptly. Conventional electroencephalogram (EEG) interpretation requires trained neurologists and is prone to subjectivity, motivating the need for automated and reliable computational frameworks. This study presents a hybrid EEG/Non-EEG machine-learning architecture employing a Kolmogorov–Arnold Network (KAN) for automatic seizure detection. Raw multi-channel EEG recordings were systematically gathered and annotated by three independent clinicians to establish binary seizure and non-seizure classifications. The dataset underwent structured pre-processing, segmentation into fixed-length time windows and conversion into a spectral feature space defined by the relative power of canonical EEG frequency bands delta, theta, alpha, beta and gamma. A customized Kolmogorov–Arnold Network (KAN), implemented in Python without reliance on conventional deep-learning libraries, was trained to associate these frequency domain features with seizure states. Weighted loss optimization was employed to counter class imbalance during training. Comparative assessment against traditional models such as SVM, Random Forest and feed forward neural networks demonstrated the superior performance of the proposed KAN model in terms of predictive accuracy, precision and recall. The findings highlight the KAN’s potential for interpretable, low-complexity and resource-efficient seizure detection, bridging the gap between computational intelligence and clinical usability.