Abstracts

Rationale: Epilepsy is a prevalent throughout the world and is often untreated because of lack of detection.  Our goal is to develop a robust seizure detection and localization algorithm that can be implemented in a wearable device allowing for remote monitoring and reliable reporting. It has long been known that neuronal activity in the brain is circumscribed and separated by an inhibitory network of GABAnergic neurons. During seizures these inhibitory networks are compromised leading to synchronicity in different parts of the brain. Our seizure detection and localization approach uses the reflection of this synchronicity in the EEG to declare the presence of a seizure and to provide localization of the seizure.  Methods: Treating the EEG leads as nodes in a network, we define the distance between the nodes as a function of the similarity between signals at the nodes; the more similar the signals the less the distance between the nodes.  We have defined two different measures of similarity, and hence distance. One measure is based on a normalized spectral difference between the signals in the Theta-Alpha (4-14 Hz) band, while the other is based on the correlation between signals at different nodes in the High-Gamma (80-150 Hz) frequency band. When the distance between two nodes falls below a threshold we declare the nodes to be connected. We have observed that the onset of a seizure will result in a sudden change in the connection status with both normalized spectral powers and distance measures. The connection status based on spectral similarity is more time sensitive while the connection status based on correlation is more location sensitive.  Therefore, during periods of high spectral activity we use the spectral similarity based connection to identify the start time of a seizure.  The correlation based connection information is used to refine the start time, identify the location of focal seizures and the identify the stop time of the seizure.  Finally, we use a low dimensional representation of the networks to validate the extracted parameters.  Each of these steps is computationally inexpensive and we plan to implement the system in a wearable device.  We tested the method on two different datasets.  One was a set of 18 patients ranging in age from 3.5 years to 18 years from the CHB-MIT Scalp EEG Database on Physionet and the other was a locally generated dataset on a single patient.  In total there were 580 hours of recordings which contained 104 seizures. Results: With both datasets,  our algorithm resulted in a median detection latency of 2 secs and an average detection latency of 8.46 seconds. This compares favorably with published algorithms operating on the same dataset which are either patient specific or seizure specific; the proposed algorithm is neither patient specific nor seizure specific.  The average false detection rate for the proposed algorithm was 0.398/hour. For seizures with known localizations the algorithm produced results which perfectly matched the actual focus of the seizure. Conclusions: A distance paradigm which defines a network connection has been shown to allow for the development of computationally efficient and accurate seizure detection and localization.  The algorithm is simple enough to be efficiently implemented. This computational approach is a novel approach for seizure detection and localization. Funding: NA