Brain Data Mining for Epileptic Seizure-Detection

Humans are suffering from various neurological disorders, due to the sedentary life style with stress and anxiety. Epilepsy is one of them, which is the second most prevalent neurological disorder after brain stroke. It has affected 65 million people, which are approximately 1% of
global population, with 80% of the patients being from developing countries. At every moment, a human brain generates signals, which communicate with each other. Once this communication breaks down or any abnormal events occur, this causes a seizure. As such, a person starts to behave abnormally for a few moments. It triggers a sudden uncontrolled disruption to the normal electrical activity of the brain and affects a particular point or the whole region of the brain. Different monitoring devices are available to monitor this unusual behaviour such as EEG and ECoG. Both of these are popularly used to reveal these brain signals. Though being detected by these monitoring devices, some challenges still exist such as accurate seizure detection, quick seizure detection, and localization. If the duration of these recordings are long, then the chances of missing seizure cases are very high.
In order to solve these challenges, we took advantage of techniques which is dominant and widely used in different domains including health. This research work is application oriented and is intended to solve the aforementioned challenges. Firstly, we process raw brain signals
(ECoG and EEG) data by exploring their statistical features. This is because signal data provides different statistical features, thereby, it helps in seizure detection.
Secondly, after examining the multiple classifiers performance, we have found the suitable and best classifier (i.e., decision forest), which is capable of detecting seizures from EEG and ECoG data sets. Third, we propose a novel approach by using a decision forest (i.e. an ensemble
of carefully built decision trees) for quick seizure detection without compromising accuracy. This allows quick seizure detection, which in turn result in faster remedial action. With regard to knowledge discovery, it can also be used for seizure localization – to identify the region of a patient’s brain that is mostly affected by seizure and also explores electrodes/channels which
are actively participating in seizure detection.
Finally, if the patient is onset and the duration of EEG recordings are long, then the chances of missing the seizure cases are very high, due to imbalance distribution of the class values in the data set. In order to solve this, we have proposed an approach by penalizing the cost of false
negative (CFN) corresponding to the duration of EEG recording for detecting the true seizure cases with high recall metrics.