A Deep Learning Approach in Predicting Seizure Type in Epileptic Patients Using EEG Signals

Epilepsy is a chronic noncommunicable disease and is the fourth most common neurological disease following migraine, stroke, and Alzheimer’s. This neurological disorder affects around three million people in the U.S. alone and 50 million people worldwide [1]. Physicians typically analyze electroencephalogram (EEG) signals to identify epilepsy; however, this process is time-consuming, resource intensive, and can be prone to human error. Rather than using traditional diagnostic processes, a computational method to detect and classify seizure types from EEG signals would lead to faster clinical decisions [2]. My research focuses on deep learning methods to categorize various EEG segments into different seizure types. I studied the ability of two different deep learning models, Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) networks. I used a public dataset that consisted of multi-channel epileptic EEG time series data from 6 epileptic patients with a total of 35 recorded seizures [3]. The accuracy of the CNN model was $\mathbf{9 1 \%}$, and the accuracy of the LSTM network was $\mathbf{8 2 \%}$. The ROC curves and the AUC values for the CNN, range from 0.97 to 1.00, for the different classes, and range from 0.79 to 0.94 for the LSTM network. In general, my results show that seizure type can be predicted and classified computationally and can be used in hospital workflow systems to aid healthcare professional. This computational method can be applied clinically by integrating predictive models into hospital systems, aiding physicians to reduce the diagnostic workload and ultimately improve patient outcomes. Furthermore, this method can be implemented for forecasting seizure occurrence to help physicians with care management.