Application of Machine Learning Classifiers for Accurate Detection of Epileptic Seizures from Electroencephalography Data

Abstract

Epilepsy is a neurological disorder which affects the electrical activity of the brain resulting in seizures with various intensities. The detection of epileptic seizure is of utmost importance to prevent any serious injuries or death to patients. Applications of Human Activity Recognition (HAR) in healthcare for continuous monitoring of Activities of Daily Livings (ADLs) can assist in the detection of abnormalities that may indicate the prevalence of neurological disorders such as Alzheimer’s, stroke, and epileptic seizures etc. Wearable sensors such as accelerometer embedded in smart phones and EEG headsets etc. can be used for continuous monitoring of ADLs. This research aims to detect seizures and neuroelectric abnormalities in epileptic patient using wearable EEG headset Neurosky MindWave. The data files are recorded in an un-constraint environment from healthy volunteers and epileptic patients. These data files are pre-processed using time and frequency domain analysis and machine learning classifiers are applied for classification of activities into stationary, light ambulatory, intense ambulatory, and abnormal activities. The proposed method is then compared against constraint environment EEG dataset from Bonn University and dataset collected from accelerometer-based application ‘MyNeuroHealth’. It is evident from the research that portable EEG headset may be employed to detect abnormalities which may lead to a seizure whereas the system implemented by ACM is restricted to detect motor components only. The results show that SVM classifier can detect ADLs and seizure with a reasonable accuracy of 96.7% in an un-constraint environment where Random Forest performs better for classification of states in a constraint environment with 50% greater accuracy than SVM classifier. Furthermore, the proposed and implemented EEG-based system can detect ADLs and seizure with 3% better accuracy as compared to accelerometer-based system.