SUBJECT-ORIENTED EPILEPTIC SEIZURE DETECTION USING SMOTE AND DEEP LEARNING ON SINGLE-CHANNEL EEG FROM THE CHB-MIT DATASET

Abstract

Epilepsy is a common neurological condition that affects millions of individuals, causing significant disruption in their lives due to its unpredictable and recurrent seizures. Seizures are unpredictable, which puts personal safety at serious risk and lowers general wellbeing.  Therefore, accurate seizure detection is essential for improved illness treatment; yet, existing approaches frequently lack effectiveness, customization, and convenience of implementation.  With the goal of improving clinical accuracy and practicality, this study explores a subject-specific method of seizure identification utilizing EEG data.  Through the use of the CHB-MIT Scalp EEG dataset, our technique is tailored for each person.  For each patient, we used data from a single EEG channel to guarantee a simplified architecture appropriate for real-world use.  EEG data were separated into eight-second segments that overlapped and were tagged with whether seizure activity was present or not. Only the training dataset was subjected to the Synthetic Minority Over-sampling Technique in order to solve the issue of data imbalance between seizure and non-seizure events.  To eliminate the need for manual feature extraction, a one-dimensional Convolutional Neural Network was developed to automatically recognize and categorize relevant characteristics from these segments.  The feasibility and reliability of seizure detection are enhanced by this customized modeling technique.  Compared to generalized approaches, more consistent performance is obtained by customizing the model to each patient's unique neural patterns.  This approach is a significant advancement in developing patient-centered, easily accessible solutions for epilepsy monitoring because of its emphasis on customization and computational efficiency.