Differential operators can detect significant changes in signals. This has been utilized to enhance the contrast of the seizure signatures in depth EEG or ECoG. We have actually taken normalized exponential of absolute value of single or double derivative of epileptic ECoG. Variance operation has been performed to automatically detect seizures. A novel method for determining the duration of seizure has also been proposed. Since all operations take only linear time, the whole method is extremely fast. Seven novel parameters have been introduced whose patient specific thresholding brings down the rate of false detection to a bare minimum. Results of implementation on the ECoG data of four epileptic patients have been reported with an ROC curve analysis. High value of the area under the ROC curve indicates excellent detection performance.
Second derivative based Laplacian operator is widely used for edge detection in an image [1]. An edge can be characterized by an abrupt change in intensity indicating the boundary between two regions of an image [2]. We have applied the same logic in this letter to detect the boundary between seizure and nonseizure in ECoG signals of epileptic patients which indicates a seizure onset or offset (for automatic seizure detection see [3] - [5]). Earlier first and second derivative of neonatal sleep EEG were used for feature extraction in order to automatically detect the sleep stages [6]. First and second derivative of EEG were also used to extract time domain features for automatic seizure detection in [7].
In the next section we will describe the method. In section 3 data acquisition will be described.
In section 4 will contain the results of implementation on depth EEG or ECoG of four epileptic patients. We will use EEG and ECoG interchangeably throughout this letter. The last section contains some concluding remarks.
In this letter we will be dealing with digital signals only. Derivative
, where 2 D is the second derivative, . is absolute value and w is a positive
acts as a spike enhancement filter with respect to the back ground EEG (spike enhancement through appropriate filter for the detection purpose has also been accomplished in [8]).
too acts the same way, where D denotes the first order derivative. Depending on the data set one gives better results than the other.
Success of variance in seizure detection is well established [9]. In section 4 we will see that the above filtering can significantly improve the seizure detection accuracy by variance. In order to minimize false detection patient specific threshold needs to be set for the following parameters introduced in this paper. For the detail of implementation see the MATLAB programs with elaborate documentation along with supplementary materials in the author’s website [10]. The implementation results for particular patients at any given time slot and channel are summarized in Fig. 1 (for single derivative filter or SDF) and Fig. 2 (for double derivative filter or DDF).
(a) Maximum windowed variance (B) of the filtered data (in the above sense).
(b) Maximum windowed variance of absolute value of the data (C).
( N is the offset (usually several seconds after the actual offset). E is an array consisting of windowed variance of the filtered data starting from two windows before M up to 1 N .
x is an array consisting of maximum values of E . F is another array consisting of values of E which are greater than or equal to ) max( 43 E .
And now we are in a position to say that
is a quantity whose threshold distinguishes between seizure and nonseizure EEG.
(e)
, where std stands for standard deviation.
, where v is a positive valued normalization constant. K is an array consisting of values of DE , which are greater than or equal to ) max( * 999 . 0 DE . Length of K , whose threshold distinguishes between seizure and nonseizure EEG.
(g) For seizure EEG ) max(DE must lie within an interval.
Let us mention once again that all the thresholding in the above parameters and the interval in (g) are patient specific. In this work duration of seizure has been calculated as described in (d) Freiburg, Germany [11]. In order to obtain a high signal to noise ratio (SNR), fewer artifacts and to record directly from focal areas intracranial grid, strip and depth electrodes were utilized. The ECoG data were acquired using Neurofile NT digital video EEG system (It-med, Usingen, Germany) with 128 channels, 256 Hz sampling rate, and a 16 bit analog to digital converter. In all cases the ECoG from only six sites have been analyzed. Three of them from the focal areas and the other three from out side the focal areas. See Table 1 for the patient details. A superset of the patient population has been studied in [12].
Since the data were collected over couple of years, the conditions under which the data had been collected are likely to be different from patient to patient. We have performed different preprocessing for different patients for the optimum results. We have chosen the method by trial and error. Gaussian low pass filter, with cut off frequencies either 50 or 100 Hz depending on the patient, has been used to remove muscle contraction artifacts. Montage change from common reference to bipolar has helped to suppress chewing artifacts in patient 4 to some extent. See Table 2 for the details. For patient 4 the three in-focus electrodes have been put in bipolar reference among themselves and three out-focus electrodes have been put in bipolar reference among themselves (Although intensity of seizure decreases due to subtracting one channel from another which may result in detection failure, in this particular case it helped to eliminate artifacts to a large extent while still preserving the strength of the signal, which has turned out to be suffic
This content is AI-processed based on open access ArXiv data.