Epileptic Seizure Prediction from the Scalp EEG Signals by using Random Forest Algorithm

Abstract

Epilepsy is a common chronic neurological disease of the brain and becomes the second critical neurological disease following the cerebrovascular disease. It has great significance of studying epileptic electroencephalogram (EEG) signals to reduce the frequency of seizure or even prevent epilepsy, thereby decrease the incidence of disability and death induced by epilepsy. Therefore, we proposed an epileptic EEG signals prediction algorithm by using the random forest (RF). The scalp EEG signals used here is weaker than the intracranial EEG signals but easier to collect. The prediction process had four steps: epileptic scalp EEG signals preprocessing by the discrete wavelet transform (DWT), features extraction by several linear measures in the time domain, comparison of the general model and the patient-specific model, and prediction optimization by RF with different features and sub-bands combination. It is found that patient-specific models had higher accuracy, and the coefficient of variation (CV) had a great influence on the seizure prediction. The influence of different sub-bands on the seizure prediction was also investigated. Finally, CV with all sub-bands were optimally selected, and the average accuracy of the seizure prediction was up to 99.8102%. The effectiveness of the proposed prediction algorithm was verified by numerical calculation.