Feature Engineering Coupled Machine Learning Algorithms For Epileptic Seizure Forecasting From Intracranial EEGs

Abstract

Epilepsy is one of the major neurological disorders affecting nearly 1 percentage of the global population. The major blunt is born by under developed and developing countries due to expensive treatment of epileptic conditions. Further, the lack of proper forecasting methods for an occurrence of epileptic seizures in epileptic-drug resistant patients or patients not amenable for surgery affects their psychological behaviour and restricts their daily activities. The forecasting is usually performed by human experts that leave a wide gap for human-bias and human error. Therefore, in the current work, we have evaluated the efficiency of several machine learning algorithms to automatically identify the preictal patterns corresponding to epileptic seizures from intracranial EEG signals. The robustness of the machine learning algorithms were tested after the data set was pre-processed using carefully chosen feature engineering strategies viz. denoised Fourier transforms as well as cross-correlation across electrodes in time and frequency domain. Extensive experimentations were carried out to determine the best combination of feature engineering techniques and machine learning algorithms. The best combination of feature engineering techniques and machine learning algorithm resulted in 0.7685 AUC (Area under the Receiver Operating Characteristic curve) on the random test samples. The suggested approach was fairly good at prediction of epilepsy in random samples and therefore, it can be used in epileptic seizure forecasting in patients where medication/surgery is ineffective. Eventually, our strategy reveals a robust method for brain disorders forecasting from EEGs.