A Study on Machine Learning Based Generalized Automated Seizure Detection System

Seizures generally happen unexpectedly that makes their detection very challenging. Electroencephalograms (EEG) help in capturing brain's electrical activity and provide aid to the clinicians for detecting and predicting seizures. EEG is quite complex and predicting seizures by visual inspection is quite challenging, time consuming and sometimes much expensive as every time a trained specialists is required for inspection. Due to these issues various expert and automated seizure detection systems have been developed for accurate and fast detection of seizures. Preliminary seizure detection and prediction algorithms were patient-specific, where the classifier's training and testing were done on the same person. The current work provides a study of generalized automated seizure detection system which can classify seizures and non-seizures signals. EEG data from CHB-MIT database is considered for analyzing and validating the proposed detection system. The signals are decomposed using wavelet decomposition tree and wavelet-coefficients are used to calculate different non-linear features viz. approximate and sample entropy, energy, standard deviation. These features are classified by different supervised machine learning classifiers such as Support Vector Machine (SVM), Probabilistic Neural Network (PNN), K-Nearest Neighbor (KNN) and Linear Discriminant Classifier (LDC). The classifiers performance is compared using sensitivity, specificity, accuracy. The PNN classifier provided reasonably good results as compared to the other classifiers. The studied detection system will help in fast, cheap and accurate diagnosis of seizures from brain signals.