The application of auto-regressive model in feature extraction of EEG in superquiescence

Author: Hong Zhi//Hu Zhongji
Affiliation: Zhejiang University, China [1]
Conference/Journal: 2nd Int Conf on Qigong
Date published: 1989
Other: Pages: 86 , Word Count: 487


Superquiescence is physiologically and psychologically human's special state. It's also a basic state of qigong, Yoga and Transcendental Meditation(TM). The author proposed a new way of recognizing superquienscence and common quiescence state by feature extraction of EEG. Because of high efficiency of autoregressive (AR) model's feature extraction and dynamic process monitoring. We use AR model to extract state character, and so to recognizing superquiescence and common quiescence state, and to make out the common characteristic of superquiescence.

Figure one is the electroencephalogram (EEG) computer analysis system . We use sampling low filter to eliminate the frequency mixing affection. Because of artifact in signal, it is necessary to preprocess before analysis. Results analysis include EEG power spectral analysis and discriminative analysis.

1. Signal Preprocessing

Base-Line wander and Power-Line interference artifaction are often seen in EEG. It is essential to remove the artifaction in order to reflect EEG's exact information by AR model. We have designed nonregrssive finite impulse respond (NRFIR) filter. The filter is of linear phase, low pass-band ripple high stop-band attenuation. The high efficiency of the filter is demonstrated by experiments of signal preprocessing. Using the filter ensures the reliability of feature extraction, thus ensures the reliability of discriminative results.

2. AR modelling

Xn = mathematical formula]
Ai, i=1,2,...,p, are parameters of AR model, p is model, s order, En is white noise or pre-estimated error. Considering the consistence of parameter estimating and algorithms high speed. We chose MAPPLE algorithm to get parameter estimation, and AR power spectral estimation or maximum entropy estimation is as follows:

mathematical equation]

Delta t is sampling interval, and sigma^2 sub e is square error. The model order p is not only related with signal's property, but also with the purpose of study. We determine p=9 by a lot of experiments, and because of EEG signal's stationary in short time, we chose the data length (one time sampling points) N=256, thus the short time is 2.56 seconds.

3. Discriminate analysis

Recognizing superquiescence and common quiescence states is the problem of pattern recognition. To solve the problem, the first step is to extract state character, the second is t do discriminative analysis. We get EEG feature by AR modelling, that is to say A=[Al, A2, A3, ...., Ap], we consider A the state character vector, then we use step-by-step discrimination to found discriminative function and classification criterion.

4. Experiment results

The EEG tester are divided into two groups. One is high level qigong group, the other group is composed of common people. We obtained power spectral array charts and discriminative classification of the two states. The results demonstrated that the two kinds of EEG pattern are greatly different from each other, and each pattern has the same parameter range in AR model. When superquiescence, EEG power spectral is greatly increased, and spectral peak frequency shifted to low, the maximum pest shifting is as high 4.70 HZ. Initial experiments demonstrated that quiescence degree is proportional to spectral peak shifting.

BACK