Feature fusion ensemble classification approach for epileptic seizure prediction using electroencephalographic bio-signals.

Introduction: Epilepsy is a neurological disorder in which patients experience recurrent seizures, with the frequency of occurrence more than twice a day, which highly affects a patient's life. In recent years, multiple researchers have proposed multiple machine learning and deep learning-based methods to predict the onset of seizures using electroencephalogram (EEG) signals before they occur; however, robust preprocessing to mitigate the effect of noise, channel selection to reduce dimensionality, and feature extraction remain challenges in accurate prediction. Methods: This study proposes a novel method for accurately predicting epileptic seizures. In the first step, a Butterworth filter is applied, followed by a wavelet and a Fourier transform for the denoising of EEG signals. A non-overlapping window of 15 s is selected to segment the EEG signals, and an optimal spatial filter is applied to reduce the dimensionality. Handcrafted features, including both time and frequency domains, have been extracted and concatenated with the customized one-dimensional convolutional neural network-based features to form a comprehensive feature vector. It is then fed into three classifiers, including support vector machines, random forest, and long short-term memory (LSTM) units. The output of these classifiers is then fed into the model-agnostic meta learner ensemble classifier with LSTM as the base classifier for the final prediction of interictal and preictal states. Results: The proposed methodology is trained and tested on the publicly available CHB-MIT dataset while achieving 99.34% sensitivity, 98.67% specificity, and a false positive alarm rate of 0.039. Discussion: The proposed method not only outperforms the existing methods in terms of sensitivity and specificity but is also computationally efficient, making it suitable for real-time epileptic seizure prediction systems.