Algorithms for EEG-Based Monitoring of Epileptic Seizures

Millions of people around the world suffer from epilepsy. Approximately 0.1 percent of epileptic patients die from unexpected deaths. It is of a great value if technology can provide a method to efficiently monitor the seizures and alert the caregivers to help patients. An Electroencephalography (EEG) signal is able to discover any neuron’s misfiring or excessive neural activity which can be a sign of a neurological disorder. It is proven that EEG signals are the best markers for detection and diagnosis of the epileptic seizures. Frequency domain features (like normalized in-band power spectral density) are known as most informative attributes to extract meaningful information from EEG signals. In this work, we addressed three main challenges in the area of epileptic seizure monitoring. First, we proposed a channel selection method which selects the most informative EEG channels out of full EEG channel set. We embedded high dimensional spectral features into the low dimension space to improve the accuracy seizure detection. Second, we suggested two novel imbalance learning techniques to address the problem of class imbalance in the seizure dataset. Using this approach the classification models can better get trained and learn more from seizure samples. Third, we proposed a personalized seizure prediction methodology to extract footprint of seizure and identify pre-seizure attributes based on each patient’s response that time. Using this approach, the accuracy of seizure prediction is improved since only the most informative portion of pre-seizure data is used for prediction.