Effective feature extraction techniques for increasing robustness of epilepsy prediction systems

Abstract

Epilepsy is a common neurological disease in which a patient experiences seizures. Nearly 65 Million people are affected by epilepsy worldwide. These patients are monitored and examined in hospitals using electroencephalographic (EEG) recordings. Surgery and medication are the two common ways of treatment of such patients. However, these treatments provide satisfactory results in around 70% of the patients whereas 30% patients do not get control over seizures. One of the most dangerous aspect of epilepsy is that seizures can occur all of a sudden that lead towards serious injuries, which in some cases turn fatal. Therefore, these seizures need to be predicted well before they actually happen. In-time and accurate seizure prediction can help prevent serious damage. The current state of the art epilepsy prediction system has an average sensitivity of 90.3% and average specificity of 85%. In this study, we propose deep learning methods for seizure prediction. Our method consists of preprocessing of EEG signals, feature extraction and classification. In preprocessing step, we have applied Short Term Fourier Transform with Overlapping/Non- OverlappingWindow to convert into frequency domain and used bandstop filtering to remove noise from EEG signals. Convolutional Neural network is used to extract features and classification has been done using Support Vector Machines. We have trained and tested our proposed method on Scalp EEG data of 22 subjects by CHBMIT. The proposed system achieved an average sensitivity and specificity of 92.8%, and 90.7% respectively. The proposed system performs better than the state of the art techniques in terms of sensitivity and specificity. In future, intelligent algorithms can be used to preprocess the data to further increase the Signal to Noise Ratio. The poor quality of EEG recordings is the main culprit for causing the False Positives. A hybrid approach of handcrafted and automated features can also help in better feature selection and ensemble classification methods could improve classification results.