EEG-Based Seizure Prediction Using Transfer Learning

Ziwei Wang,Dongrui Wu

2021-01-01

Abstract:As one of the most common brain diseases, epilepsy affects nearly 1% of all human beings. Patients’ central nervous systems are destroyed chronically, hardly differentiable from a normal state until epileptic seizures happen, and hence difficult to prevent. About 65 million people worldwide have epilepsy, more than the combined number of people diagnosed with Parkinson’s disease, Alzheimer’s disease and multiple sclerosis. It is practical and feasible to establish a reliable seizure prediction system. When the electroencephalogram signal is predicted to be preictal, the system sends an alarm to remind patients, family members or doctors to take drug treatments, electrical stimulation or physical measures to prevent seizures. We introduce three transfer learning algorithms, including TCA, JDA and EasyTL, together with two ensemble learning algorithms, RF and GDBT, to predict seizures, constructing a seizure prediction system framework based on transfer learning algorithms. Three experiments are conducted consecutively on a public dataset: 1) single patient seizure prediction experiment reaches prediction of 95.18%; 2) cross-patient seizure prediction without transfer learning results in prediction accuracy of 32.71%. Compared with Experiment I, Experiment II confirmed that the individual differences of EEG among patients would lead to a significant decline in prediction performance; 3) cross-patient seizure prediction with transfer learning. After using transfer learning algorithms, the prediction accuracy was improved from 32.71% to 57.29%, demonstrating the effectiveness of our proposed approach and confirming the strength of transfer learning methods cross patient seizure prediction task.