oa Enhanced Epileptic Seizure Onset Detection via the Combination of Channel Selection and Feature Enhancement

Background & Objectives

Epilepsy is a neurological disorder that is associated with repeated episodes of seizures. In epilepsy, the normal pattern of neural activity is disturbed, causing the patient to experience various symptoms ranging from staring blanking for a few seconds to long periods of vigorous convulsions and unconsciousness. In many patients, the injuries they endure are a direct result of the confusion, loss of muscle control, and unconsciousness caused by the seizure. These injuries include fractures, head injuries, and burns. In an attempt to mitigate such risks, extensive research has been dedicated to developing a device that can detect or predict the onset of seizure episodes. The clinical behavior of an epileptic seizure is preceded and then accompanied by electroencephalographic alterations. As a result, electroencephalography (EEG) is the most common tool to measure these alterations. EEG measures the voltage fluctuations resulting from ionic current flows within the neurons of the brain. Research has demonstrated that epileptic seizures are caused by disturbances in the electrical activity between neurons in the brain. In the healthy brain, neurons fire in an asynchronous manner, relaying messages from one neural network to the other. However, in the epileptic brain, the complex interactions between neuronal networks are characterized by the evolution of synchronization between them. Excessive neuronal synchronization leads to a hyper-synchronous state that triggers the onset of a seizure. Extensive research has been dedicated to the detection of the earliest signs of electrographic changes associated with a seizure using either scalp or inter-cranial EEG. A device that has the ability to detect the electrographic onset of a seizure (seizure onset detector) will enable epileptic patients to lead a more normal and secure life, and will help them to avoid injuries due to the sudden nature of the seizure. At its most basic form, a seizure onset detector (SOD) is comprised of two major components, the feature extraction unit and the classification unit. In the feature extraction unit, relevant features are extracted from the multi-channel EEG and are fed into a classification unit where the features are classified as seizure or non-seizure in nature.

Methods

In this research, we present a novel patient-specific epileptic seizure onset detector using scalp EEG where we aim to exploit the benefits of integrating EEG channel selection and feature enhancement to improve the SOD's performance. Hence, we propose a detector architecture that is composed of five stages, namely, EEG channel selection, feature enhancement, spatial averaging, feature extraction, and classification. Assuming the collected scalp-EEG data contains M-channels, the channel selection stage chooses the N EEG channels that contain the most relevant electrographic seizure information. This stage is important because it reduces the detector's computational burden and omits the need to evaluate channels that have invaluable information, which may deterioted the detector's performance. The main goal of the feature enhancement stage is to emphasize the seizure EEG relative to the non-seizure EEG (background EEG). For this, a differentiation and exponentiation approach is adopted. Following the feature enhancement stage, the next stage is spatial averaging. In this stage, the N selected channels with enhanced features are averaged so that a single EEG vector is obtained. The main idea of this stage is to reduce the number of features extracted from the EEG channels to further decrease the computational complexity. The onset of a seizure is usually associated with rhythmic activity composed of multiple frequency components. Therefore, to utilize the information contained in the EEG, features are extracted from each frequency sub-band separately. A multi-resolution wavelet decomposition is used to extract relevant frequency components from the EEG. From this point, the energy from each frequency EEG sub-component is calculated as the EEG feature. The last stage of our detection is the classification and detection stage. A support vector machine (SVM) is used to classify EEG epochs as seizure or non-seizure.

Results

The data used to evaluate the proposed detector is from a publicly available database consisting of EEG recordings from pediatric subjects with intractable seizures, collected at the Children's Hospital Boston. The subjects have been monitored for up to several days following withdrawal of anti-seizure medication in order to characterize their seizures and assess their candidacy for surgical intervention. The proposed detector was tested on six of these patients, where a leave-one-out cross-validation testing scheme is adopted for each patient. In the leave-one-out cross-validation testing scheme, the SVM is given a training set that includes the seizure and non-seizure epochs from all but one of the subject's recordings. The detector then attempts to detect the seizure from the excluded recording using the learned knowledge from the training set. This process is repeated until all recordings have been tested. To test the effectiveness of our proposed detector, the energy difference between pre-seizure and seizure states are calculated for a detector that only implements spatial averaging, a detector that implements a feature enhancement stage with no channel selection, and our proposed detector. It is found that the energy difference when feature enhancement is used is higher than when no feature enhancement is implemented in the detector. The energy difference is further enhanced for the case when the detector implements both feature enhancement and channel selection. The performance of the detector is compared to a benchmark detector that uses only feature enhancement. Our proposed detector achieves a detection latency of 4 seconds, which is closely aligned with the electrographic seizure onset time that an expert has visually determined. Furthermore, the proposed detector achieves a sensitivity of 87.5% whereas the benchmark detector achieves a sensitivity level of 80.6%.

Conclusion

Our research proposes a novel architecture for an epileptic seizure onset detector. The combination of the channel selection and feature enhancement stages has led to an improved detection performance. An increase in the energy difference between seizure and pre-seizure states is observed when the proposed detection is implemented, versus implementing the detection system without any form of channel selection. The proposed system also performed better in terms of false alarm rate and sensitivity.