Using machine learning to predict epileptic seizures from EEG data

Abstract

With the recent progress in Machine Learning (ML) and Artificial Intelligence (AI), researchers aim to apply techniques for improving and automating certain facets of clinical practice. One of the more intriguing and compelling applications of modern computing in a healthcare context is the early detection and prediction of life threatening events. In the case of epilepsy, the prediction of seizure onsets would allow patients to appropriately prepare for such recurrent episodes of convulsion, which in turn improves their quality of life. Albeit seizures are preventable by specific medication and therapies, it is common for patients to suffer from intractable seizures, which is the result of drug-resistant epilepsy. The prediction of seizure onsets would allow patients some relief in knowing when to be prepared and when to avoid dangerous activities such as driving. This Bachelor’s dissertation presents a review of the performance of a set of supervised machine learning methods for the task of seizure prediction. The study involves using a dataset that includes non-invasive scalp Electroencephalography (EEG) signals, which are brain electrophysiological readings that did not involve surgery. Subsequent to data pre-processing, statistical and wavelet features from the signals were extracted, and the results obtained from K-Nearest Neighbour (KNN), Support Vector Machines (SVM), and an Ensemble Classifier are compared. Results are reported on the CHB-MIT dataset, which includes 192 seizure readings from 22 patients suffering from intractable seizures. The study shows that the three methods perform similar, although the Ensemble Classifier achieves a higher specificity, sensitivity and accuracy.