Article Preview
TopIntroduction
For the people with epilepsy disorder, the moment the seizure occurs will be very dangerous and challenging. Nearly 1% of the population suffers from epilepsy even today. The seizure begins with a sudden unexpected storm of electrical signal in some region of the brain. The signal may be emitted in one or two regions of the brain and may also spread to other regions. This causes convulsions and loss of consciousness in case if the seizure is very severe. Such situations become very terrible for the patient and people around him to handle. This disrupts the daily activity and lifestyle of the person suffering from epilepsy. Though neurologists provide best available anti-epileptic drugs to the patients, patients continue to suffer seizure at times. The fear of unexpected seizures makes patients to consume excessive medicines and suffer side effects. Therefore, forecasting epileptic seizures is a good attempt which can improve the treatment and in turn the life of the epilepsy patients to a far extent (Brinkmann, 2015) (Javier, 2017). A precise prediction could help them to avoid risky activities, handle the anxiety levels and avoid unnecessary medications. Alternate method followed by the neurologists is to go with a surgery and remove the part of the tissue from the brain which is found to be causing seizures. Still one more method identified was to implant a neurostimulator device. These devices keep track of the electrical pulses in the nervous system and prevent the sudden emission of high energy signals. This prevents sudden unexpected seizures. But too many electrical pulses could not be controlled by the device and hence this cannot be the best solution. Therefore, it is necessary to study the electrical patterns of the seizures in the brain in order to predict them.
Many different algorithms, machine learning approaches are used so far by the researchers to perform forecasting of EEG signals (Nguyen, 2020) (Iwok, 2016) (Kulkarni, 2020). Deep learning technique: long short-term memory is a best technique for the analysis of time series data sets. Since EEG recordings are collected along with the time steps continuously, LSTM is the best method for analysis.
In this paper we perform EEG forecasting in two parts; once using univariate time series and second time using multivariate time series (Kim, 2013). In case of univariate time series forecasting, the prediction of the next time steps is performed only based on the earlier time steps of a single attribute. The response variable i.e. the target variable is influenced by only one factor (Aqsa Shakeel, 2020). Whereas, in case of multivariate time series forecasting; the prediction of the future time step value is performed based on multiple time dependent attributes (JD Peter, 2019). Each attribute will have dependency not only on the earlier time steps of its own but also on the earlier time steps of the other attributes(K J Blinowska,1991)(R. Ranjan, 2017). Thus the response variable is influenced by multiple factors. Because of this reason multivariate time series forecasting is considered to be more accurate than univariate time series.