Analysis of Intracranial Electroencephalography for Prediction of Epileptic Seizure

Abstract

Epilepsy is one of the most common and devastating neurologic diseases, which affects over 70 million people around the world. For some patients, it can be managed with antiepileptic drugs. However, 20 to 30% of them would likely get worse after the initial improvement, and some may even remain refractory to the current medicine. Predicting seizures enough in advance could allow patients or caretakers to take appropriate actions and therefore, reduce the risk of injury. This work focused on time-domain features to achieve discriminative information about EEG signals and at a lower CPU cost. Thus, Kruskal Wallis, a non-parametric approach, was found to perform better than other approaches due to its less resource consumption strategy while maintaining the highest Matthews’s Correlation Coefficient (MCC) score. The performance of Kruskal Wallis may suggest considering the importance of univariate features like complexity and interquartile ratio along with auto regressive model parameters and maximum cross-correlation. Furthermore, it has been demonstrated that dividing EEG to sub-bands will provide more discriminative information and that a 2-second window length provides the highest MCC score while dealing with the non-stationary behaviour of EEG signal for this window length. Consequently, the obtained results from 2-s window were fed to a binary (Extreme Gradient Boosting) classifier and the posterior probability of the test data extracted. The probabilistic framework requires the mean and maximum probability of the non-seizure and the seizure occurrence period segments. Once all these parameters were set for each patient, the medical decision maker can send alarm based on well-defined thresholds. While finding a unique model for all patients is really challenging, our modelling results demonstrated that the proposed algorithm can be an efficient tool for reliable and clinically relevant seizure forecasting. Using iEEG signals, the proposed algorithm is capable of forecasting seizures, informing a patient about 75 minutes before a seizure would occur, a period large enough for patients to take practical actions to minimize the potential impacts of the seizure. The proposed tool aims to be implemented in a low power portable device by considering few simple time-domain features for a specific sub-band. It should be noted that there are not many publications investigating frontal lobe epilepsy, making the findings in this work promising results.