Pre-Epileptic Seizure Detection Using Machine Learning

Abstract

Epilepsy is a chronic neurological disorder that affects millions worldwide, characterized by recurrent seizures caused by abnormal electrical activity in the brain. Pre-seizures refer to abnormal patterns or symptoms that occur minutes before a seizure and are critical to detect for timely interventions. As seizures can strike during daily activities like driving, accurate prediction holds significant clinical and practical value. In this study, we propose a robust seizure forecasting framework using EEG recordings to predict the pre-seizure state. The system includes EEG signal preprocessing, feature extraction, and classification, focusing on transitions from normal to interictal (between seizures) and ictal (during seizures) states. Key features such as mean, variance, and delta power are extracted using statistical methods, and classification is achieved through a decision boundary-based system. The novelty of our approach lies in the integration of statistical feature extraction with neural network-based classification and the use of Frequency-Dependent Synaptic Plasticity (FDSP), a biologically inspired mechanism that adapts to frequency variations in EEG signals. This enhances the system's ability to detect seizure onset dynamically and accurately. Classification thresholds are determined using normalized differences and the McCulloch approach, refining decision boundaries for better state identification. Compared to traditional models, our method achieves higher prediction accuracy, sensitivity, and robustness. It offers smoother transitions between states and serves as an effective early-warning system, the model is well-suited for deployment in low-resource settings, making it both technically efficient and socially impactful.