A Hybrid Deep Learning Architecture for detection of epileptic seizure state using brain signals

Abstract

Epilepsy is a chronic neurological disorder that occurs due to irregular brain activ- ities. Epileptic seizures require the continuous attention of the patients and caregivers. Epilepsy is generally treated with anti-epileptic drugs (AEDs), with many side effects like unsteadiness, poor concentration, sleepiness, double vision, vomiting, and tremor. These side effects affect the quality of life of the patient. Electroencephalography (EEG) is widely used to treat epilepsy as it captures the brain s electrical activities. An auto- mated approach to detect the seizure state from EEG recordings is highly desirable as the manual approach is tedious, time-consuming, and prone to errors. EEG signals are non-stationary and noisy due to the muscle activities while recording the brain signals, varying from person to person and within a person with age. EEG signal recordings also pose the issue of high-class imbalance as the seizure events are very few compared to non-seizure events in several hours of recordings. newlineResearch in this area is flourishing, and many seizure detection devices are available. However, higher sensitivity and a lower false-positive rate are desirable for clinical use. This research aims to develop an epileptic seizure state detection algorithm that will assist neurophysiologists in making quick and accurate seizure diagnoses. This research work first evaluates the performance of eight classical machine learning algorithms using the feature extraction method, which includes: k-nearest neighbor (k-NN), decision tree (DT), Gaussian Naive Bayes (GNB), multi-layer perceptron (MLP), quadratic discriminant analysis (QDA), support vector machine with Gaussian and polynomial kernels (SVM-G, SVM-P), and random forest (RF). The benchmark seizure classification dataset of Bonn University was used to perform the experiments. The results demonstrated that GNB, SVM-G, and RF perform better if given noise-free, class-balanced, and small-scale datasets.