Epileptic seizure prediction with machine learning on EEG data

Epileptic seizure prediction with machine learning on EEG data

Epilepsy affects over 50 million people globally, posing a significant challenge due to the unpredictability of seizures, which impacts patients' quality of life. Predicting epileptic seizures in advance can improve living standards through timely interventions and risk reduction. However, accurate seizure prediction remains unsolved. This research aims to enhance seizure prediction accuracy using machine learning methods applied to epileptic electroencephalography (EEG) signals. Key methods developed include personalized channel selection, efficient feature extraction, and accurate classification models. Three predictive models were developed, each showing a remarkable performance in terms of accuracy, sensitivity, and specificity, leading to notably enhanced epileptic seizure prediction rates. The first model employs a permutation entropy-based personalized channel selection, significantly improving accuracy but requiring careful consideration of factors that influence the channel selection. The optimal channels for predicting seizures may vary across different stages of the condition. Hence, the second model employs a personalized classification for the entire channels from each patient, emphasizing the superiority of Synchroextracting Transform (SET) over the popular short-time Fourier transform for accurately extracting information. SET, when combined with a one-dimensional convolutional neural network (1D-CNN), achieves a 100% accuracy, sensitivity, and specificity for the Bonn University database (82800 datapoints), surpassing a multilayer perceptron with a quicker computational speed. Meanwhile, when considering real-time monitoring of epileptic EEG, CNNs may not be suitable due to their computational costs and substantial memory requirements. Therefore, in the third model, a sparse representation combined with SET and basic traditional machine learning techniques like k-nearest neighbors are adopted. This approach has also been proven to be notably effective, with a 100% accuracy on the Bonn University database for seizure prediction. The three models developed in this research address the challenges arising from individual variability in brain functions and the high-dimensional nature of EEG data for epileptic seizure prediction. Future research should focus on optimizing these models for specific real-time EEG monitoring systems and conditions.

File reproduced in accordance with the copyright policy of the publisher/author.