Epileptic seizure detection and prediction using stacked bidirectional long short term memory

Abstract

Epilepsy is a not a disease but a neurological disorder. But people with epilepsy cannot lead a social life like others. Proper diagnosis and advance prediction of epileptic seizures definitely improves the life of epilepsy patients. In this paper an effort is made to develop a seizure detection and prediction method using stacked bidirectional long short term memory technique. This is the most suitable technique for the analysis of time series datasets as it overcomes the vanishing gradient problem identified in recurrent neural network. The dataset for detection and prediction experiments was taken from Bonn University. Our model could perform the seizure detection with the highest accuracy of 99.08% with 98% precision, 99.5% recall and ROC AUC: 0.984346. A binary classification method with AUC more than 0.9 is considered to be outstanding. Seizure prediction was conducted using the same dataset by classifying preictal states of EEG from interictal and ictal states. For the case of prediction our model could identify preictal states with the overall sensitivity: 89.21% and false prediction rate: 0.06. In future the model could be used to program the wearable devices like wrist watch which can be used by epileptic patients for seizure prediction. The device could be programmed to fire alarm during the detection of the preictal EEG signal before the onset of the seizure.