Spark-Based Deep Learning for Recognition of Epileptic Seizures With Multimodal Brain Waves

Introduction

The brain is the body's control center and part of the nervous system, including an extensive network of neurons and the spinal cord. Brain disorders can affect our body functionalities, including a person's memory, personality, and sensation. The conditions that cause such disorders can be 1) Illness, 2) Genetics, or 3) Traumatic injury. An Epileptic seizure is one such brain disorder caused by the above conditions. Around 1% of the world's population gets affected by epileptic seizures. Of this 1%, two-thirds get cured of antiepileptic drugs. Another 7-8% get cured by epileptic surgery, but at the same time, it is challenging to diagnose seizures in patients. Therefore, around 25% of patients with epilepsy seizures are severe to discover and thus are uncontrollable (Litt & Echauz, 2002). Due to seizure activity, anxiety becomes part of a patient's life and affects them in their routine activities. Seizure activity is so harmful that it causes irreversible changes in the brain, which sometimes remain unprovoked. Consequently, this has been a significant concern among clinical researchers and a substantial area of interest in epileptic disorder (Sun et al., 2018).

Recording a patient's EEG could help the clinician to detect seizure activity. Before, EEG could only be performed in a hospitalized environment, but portable EEG systems were introduced due to the technological advancement of pervasive sensors (Waterhouse, 2003). Nevertheless, it becomes difficult to study the EEG recording of an epileptic patient due to its size and complexity, which requires highly trained professionals at the clinical level. It also becomes an arduous and time-consuming task to detect seizure activity from the EEG dataset (Mormann et al., 2006). The complexity of the EEG recording dataset has inspired researchers to find a procedure that can detect epileptic seizures automatically, which can match the precision of any clinical professional. These machine learning algorithms find patterns between the attributes of the EEG dataset and perform iterations in a closed loop to make the model more accurate toward precision.

Literature survey- This section outlines various previous approaches for seizure activity prediction in an epilepsy patient. All of these technologies have impacted clinical research in multiple ways. After critically analyzing recent research on seizure activity prediction that has used machine learning algorithms, it can be said that modern technologies have been introduced to the world of neuroscience to improve the clinical sector. Brinkmann et al. (2015) showed that the preictal state of epilepsy was diagnosed from continuous ECoG in dogs with the help of Support Vector Machines (SVM). Zhang et al. (2016) showed the extension of this technique by introducing the pre-processing of ECoG data and then having it classified from the Kalman filter, which improved the input data for the SVM linear classifier. Wang et al. (2015) proposed the processing time by decreasing the redundancy in the dataset. Doing further refining by using the global features of datasets were easily classified. A specialized seizure detection with EEG scalp was introduced by Vidyaratne et al. (2016), which was automated with the help of a novel Deep Recurrent Neural Network (DRNN) and was patient-specific. For seizure detection, Turner et al. (2014) worked with high resolution and multichannel EEG data using the Deep Belief Network (DBN). Yuan et al. (2017) identify abnormalities for seizure detection using multichannel epileptic EEG signals.