Feature Selection for Personalized EEG-Based Seizure Monitoring




Journal Title

Journal ISSN

Volume Title




Epilepsy refers to a group of neurological disorders characterized by neurons in brain that misfire and cause recurrent seizures. Epilepsy affects millions of people around the world. Epileptic patients who have intractable seizures may suffer from injuries or even sudden death. It is of great value and significance to provide continuous, reliable and timely seizure monitoring for epileptic patients. Scalp electroencephalogram (EEG) signal is widely considered the golden marker of epileptic seizures. A low-cost, wearable EEG-based seizure monitoring platform using machine learning technology will not only improve patients’ quality of lives, but also assist physicians for clinical diagnosis. Clinical EEG signals are collected from multiple channels (e.g., 10 to 30 electrodes), and usually have highly complex patterns that vary among different patients. To have a wearable monitoring device for a patient, it is important to determine a limited number of seizure-relevant channels (e.g., no more than 5) and select a few discriminative features of that subject. This thesis presents two methodologies on personalized feature selection methodology to enhance EEG classification performances for wearable seizure monitoring. First, a feature ranking approach using F-statistic value is presented to select discriminative features from fusion feature set consisting of spectral power and entropy features. The selection results of top-ranked features are visualized for personalized analysis. Second, to design a wearable EEG monitoring platform, a two-step feature selection strategy is proposed. Step 1 uses linear discriminant analysis (LDA) to find seizure-indicative channels. Step 2 employs the least absolute shrinkage and selection operator (LASSO) method to select a subset of spectral features. What’s more, a personalization scheme is proposed to choose the best selection parameters of each subject. The proposed methods have been evaluated on public domain CHB-MIT EEG database, and the results are better or comparable to state-of-art works reported in literature.



Electroencephalography, Epilepsy, Spasms, Machine learning