Dimensionality of IED and Non-IED Waveforms.
Abstract number :
1.105
Submission category :
Year :
2001
Submission ID :
1957
Source :
www.aesnet.org
Presentation date :
12/1/2001 12:00:00 AM
Published date :
Dec 1, 2001, 06:00 AM
Authors :
M.D. Bej, M.D., Dept. of Neurology, Cleveland Clinic Foundation, Cleveland, OH; R.C. Burgess, M.D., Ph.D., Dept. of Neurology, Cleveland Clinic Foundation, Cleveland, OH
RATIONALE: To determine the dimensionality of interictal epileptiform discharges (IEDs) and non-IED waveforms by using dimension reduction via artificial neural network (ANN) pattern matching.
METHODS: A 5-layer ANN was trained on raw EEG as previously described, with desired output values set to be identical with the input values. Three hidden layers were employed, with the middle layer containing very few (n=3, 4, 5) neurons. The network was trained to mean standard error (MSE) stability, ca. 10,000 iterations.
RESULTS: This method has previously been shown to separate IEDs from nonIEDs reliably at n=3. Separation of IEDs and nonIEDs was also achieved with n=4 and 5. The goodness of separation improved as middle-layer nodes were added, indicating that the EEG was better characterized with greater numbers of such nodes.
CONCLUSIONS: For effective EEG feature extraction, the number of expected features should be known and limited. Extraction of more features than what is warranted leads to excessive work and features that are not independent from each other. This method provides a mechanism for determining the [ssquote]practical[ssquote] dimensionality of any signal. At the same time it provides a method for separating IEDs and nonIEDs at an arbitrary level of precision, as well as at a user-determinable cost/time ratio.