Spatial Temporal Cross Modality Deep Learning for Multi-Electrode EEG-Based Epilepsy Detection

Abstract

The broad spectrum of neurobiological disorders caused by abnormal electrical activity in brain results in recurrent seizures called as epilepsy. Early prediction of seizure onset is crucial, as it provides a vital time window for timely intervention and treatment. In our previous work, we introduced a multimodality framework for epileptic seizure detection using EEG features. Though it improved the accuracy, it could not reduce the seizure detection onset time. This is due to ineffectiveness in learning correlation between the EEG signals collected from multiple electrodes. The earlier work considered cross modality only from feature representation view. It did not use the multi electrode signal correlation for seizure detection. To solve this problem, this work proposes spatial temporal cross modality learning to capture local and long range interactions between EEG signals across multiple electrodes. Graph neural network is used to learn cross modality features from multiple electrodes. The features are then classified by optimized Long Short-Term Memory (LSTM) network. Optimization of hyperparameter is achieved with a variant of elephant heard optimization (V-EHO) algorithm.