Predicting Epileptic Seizures from Intracranial EEG Using LSTM-Based Multi-task Learning

摘要

Epilepsy afflicts nearly 1% of the world's population, and is characterized by the occurrence of spontaneous seizures. It's important to make prediction before seizures, so that epileptic can prevent seizures taking place on some specific occasions to avoid suffering from great damage. The previous work in seizure prediction paid less attention to the time-series information and their performances may also restricted to the small training data. In this study, we proposed a Long Short-Term Memory (LSTM)-based multi-task learning (MTL) framework for seizure prediction. The LSTM unit was used to process the sequential data and the MTL framework was applied to perform prediction and latency regression simultaneously. We evaluated the proposed method in the American Epilepsy Society Seizure Prediction Challenge dataset and obtained an average prediction accuracy of 89.36%, which was 3.41% higher than the reported state-of-the-art. In addition, the input data and output of middle layers were visualized. The visual and experiment results demonstrated the superior performance of our proposed LSTM-MTL method for seizure prediction.