A Deep Learning Approach for Real-time Detection of Epileptic Seizures using EEG

Despite advances in neurosurgical and drug therapy procedures suggested for treating epilepsy, about 30% of patients are considered as refractory cases that remain non-responsive to treatments. In such cases, the patients experience seizures varying in severity and frequency, from one or fewer seizures per month to multiple seizures per day. This unpredictable nature of seizures and their potential threat to patients' lives calls for a sustainable solution for the long-term monitoring and real-time detection of epileptic seizures. Such a solution would allow caretakers and physicians to monitor seizures' occurrence and severity, administer care in a timely fashion and intervene to prevent harm. In this work, we propose a deep learning-based approach for the automatic detection of a large variety of epileptic seizures, in real-time, and with high accuracy and low false positives. The proposed approach consists of a sophisticated Bi-LSTM deep learning model that uses as input 19 EEG inputs, 134 wavelet features, and the arithmetic mean of the EEG data. The model was trained and tested on the most comprehensive dataset available for Epilepsy data, encompassing EEG recording for 161 Epileptic patients and 1005 labeled seizures of varying types. The obtained results are very promising and open the door for the automatic monitoring and real-time detection of seizures within less than 2 seconds, with high detection accuracy and low false negatives.