F134. Automated seizure detection using statistical CUSUM detector

Introduction Although automated seizure detection methods using intracranial EEG (iEEG) have achieved high accuracy in previous studies, they acquire many labeled datasets. Also, due to the non-stationarity nature of seizures and the inter and intra-individual variability in signal characteristics, these methods are difficult to implement prospectively in clinical practice. We propose an automated seizure detection method using a cumulative sum (CUSUM) detector that can be used online with fewer training parameters and minimal overall training without the need for labeled datasets. Methods The proposed seizure detector is composed of two main steps, feature extraction followed by detection.The features extracted are a line length (LL), relative energy (RE), coefficient of variation of amplitude (CVA), and the relative amplitude (RA).The main assumption for the extended CUSUM analysis is that the distributions corresponding to normal and seizure EEG are different. Feature vectors are calculated using windows of length N, subdivided into M segments of length n. At each point, the average of each of the M segments is calculated.Assuming that n is large enough, the central limit theorem applies and the sample mean vector of each segment follows a Gaussian distribution, which can be characterized by its mean and variance.A null hypothesis is formed that incoming data will be governed by the same distribution.During training, normal EEG data from the same subject is used to calculate the mean and variance bound distributions, representing the null hypothesis.During detection, for each incoming data segment, the log likelihood cumulative sum needs to be determined for each of these bound distributions.If the null hypothesis is rejected in any case, then a change is assumed to have occurred.Two 24-h long iEEG recordings containing 3 and 9 seizures respectively, were collected (sampling frequency of 2 kHz) from one patient, undergoing right parietal stereo-electroencephalography, (University of Pittsburgh IRB No. PRO15100311). Recordings were labeled by an expert closely familiar with the patients.For each iEEG file, the learning period was chosen to be the first seizure-free hour.The window length used for learning is 2.5 min. Results The CUSUM detector managed to detect the three labeled seizures of the first iEEG recording with a Good Detection Rate of 100%.While the Good Detection Rate results of the second iEEG recording are (LL = 78%, RE = 78%, RA = 88% and CVA = 78%). The number of false detections per hour results for the first EEG recording as follows (LL = 1.6, RE = 1.3, RA = 1.4 and CVA = 1.5). While for the second recording (LL = 1.3, RE = 1, RA = 1.2 and CVA = 1.25). Conclusion Seizure detection using the extended CUSUM test appears to be a promising technique for clinical monitoring purposes.This novel method for automated seizure detection using iEEG is capable of differentiating seizures from normal activity, without the need for highly customizable parameters or previously labeled data.The method also could be applied toward scalp EEG data.