Epileptic Seizure Detection From EEG Signals Using Bagged Ensemble Approach

Introduction

Epilepsy is a neurological condition in which nerve cell activity of the brain becomes disrupted. Epilepsy is one of the most common neurological diseases, and it takes fourth in terms of incidence after migraine, stroke, and Alzheimer’s. Nearly fifty million people suffer from this disorder all over the world. An epileptic seizure is a period of sudden recurrent and transient disturbances of perception or behavior resulting from the excessive synchronous activity of neurons in the brain. Only two third of epileptic patients can be cured using medications. The remaining third develop drug-resistant epilepsy and will go through surgery where seizure onset zone (SOZ) is removed. There are many investigations, and precautions are taken before performing surgery, out of which Electroencephalogram (EEG) monitoring is first done for capturing electric pulses which in turn used for finding the spot where seizure has started. Epilepsy is caused by irregular electric activities of brain and Electroencephalogram (EEG) is used to record those electric pulses. EEG data for hours is gathered and in earlier days some experts used to observe those reading and predict epilepsy. Normally the data collected is observed by 4 experts for hours long for one patient, which becomes a very big task observing for more than one people. By using machine learning techniques we can faster the prediction time for epilepsy and can obtain results in real time(Satapathy, Dehuri, and Jagadev 2017)(Vidyaratne and Iftekharuddin 2017).Alcohol use disorder detection using EEG by machine learning is performed by (Anuragi and Singh 2019). Electroencephalogram is capable of reading the electrical activity of the brain without painful and invasive. EEG is a very powerful tool with voltage range 3-100µV which is 100 times weaker than ECG (Electrocardiography) signals. To record brain activity, the electrodes connected the wires are directly attached to the scalp. Although the EEG test is useful for recording the electrical pulses, we need more expertise or skilled specialists for visualizing those signals and making more accurate inspections. There may be some previous information which can result in unpredictable behavior of present signals. Actually such information cannot be obtained directly from the recorded signal, as it can be masked by other biological signals. Therefore, the signal processing is inevitable to magnify relevant information and to extract important features from it. Generally, a process known as sub-band coding or multi-resolution signal processing which decompose given the signal to sub-bands by means of techniques like Lyapunov exponent, Fourier, Hilbert, Wavelet transforms (WTs), etc(Tsiouris et al. 2018).

Epileptic seizures can be categorized as partial or general. Partial seizures are usually produced from a localized area in the brain and may spread to other areas. Partial seizures are divided into simple and complex depending on the patient’s response during a seizure. Generalized epileptic seizure involves the entire brain and produces bilateral motor symptoms generally accompanied by loss of consciousness. Both types of epileptic seizures can occur at all ages. Two categories of abnormal activities could be found in EEG recordings of patients suffering from epilepsy: interictal which consists of abnormal signals recorded between epileptic seizures and ictal which is defined as the activity recorded during an epileptic seizure. The EEG signature of an interictal activity is occasional transient waveforms, as either isolated spikes, spike trains, sharp waves or spike-waves(Rao et al. 2016). EEG signature of an epileptic seizure is composed of a continuous discharge of polymorphic waveforms with variable amplitude and frequency, continuous spikes observed over a duration longer than the average duration of interictal abnormalities.

Different features are extracted depending on the theory that wave signals are of different amplitude and different frequencies. In this paper, we compare different machine learning models and ensemble models, before and after decomposition of the wave signal. Decomposition of wave signal is done by Discrete wavelet transform method which is the most used method. Evaluation of different models is done by deriving different evaluation metrics like accuracy, sensitivity, specificity, etc.