Pediatric Seizure prediction from EEG signals based on unsupervised learning techniques using various distance measures

Epilepsy or recurrent seizures is one of the most common non communicable neurological disorder that is prevalent in today's world population are sudden outburst of excess electrical activity of the neurons. Epilepsy can be detected from Electroencephalogram (EEG) as EEG captures and presents the electrical activity of the brain. Non-invasive EEG or scalp EEG is generally used where electrodes are placed on the scalp in order to record the brain activity. In this work a unsupervised machine learning technique is explored which is used to cluster and extract features from EEG recordings (noninvasive) to detect seizures. A patient specific approach is adopted on an open dataset (Physionet database) from where 51 seizure and 51 non seizure recordings of pediatric subjects (age ranging from lyrs to 12yrs) are considered for the related work. Unsupervised algorithm used here is the k-means algorithm to cluster the recordings into two distinct clusters of seizure and non-seizure data. When the performance of the algorithm was tested the algorithm worked with 91.43% accuracy from nearly 18, 00, 000 data taken from various subject. In future scope of work the accuracy of the method can be enhanced using appropriate features for distinctly identifying different intractable seizures according to their characteristics that are prevalent among pediatric patients.