基于AdaBoost算法的癫痫脑电信号识别

AdaBoost算法作为Boosting算法的经典算法之一, 在人脸检测和目标跟踪等领域得到了广泛应用, 但该算法也有一个缺点-退化问题. 为了解决这个问题, 通过对弱分类器进行筛选、引入平滑因子和权值修正函数三个措施对算法进行优化, 并将优化后的算法与小波包分解相结合应用到癫痫脑电信号的识别上. 结果表明, 本文算法对癫痫脑电信号的识别率为96.11%, 对正常脑电信号的识别率为99.51%, 具有较高的识别率, 为癫痫的正确诊断提供了一种可能有效的解决方案.

Recognition of epilepsy electroencephalography based on AdaBoost algorithm

Automatic recognition of epilepsy electroencephalography (EEG) signal has become a research focus because of its high efficiency, and many algorithms have been put forward to achieve it. As one of the classic algorithms of boosting algorithm, AdaBoost algorithm has been widely used in face detection and target tracking fields, but the algorithm also has a disadvantage that is its degradation. In order to solve this problem, this paper puts forward three measures to optimize the algorithm by filtering the weak classifiers whose recognition rates are low, introducing the smoothing factor and a weighted correction function. In order to verify the robustness of optimized algorithm, we choose three main parameters, i.e., the number of weak classifier, which is denoted by T; the base of logarithmic function, which is denoted by α; the threshold of weight, which is denoted by β. The experimental results of optimized AdaBoost show that it has good robustness and high recognition rate. #br#In this paper, we divide the whole process into three steps. The first step is to use the Butterworth digital low-pass filter in which the cutoff frequency of pass band is 40 Hz to filter noise whose frequency is above 40 Hz. The second step is to do feature extraction with the help of wavelet packet decomposition. The third step is to compute the sum of absolute value which are the wavelet packet coefficients of fourth layer, the wavelet package entropy and the sum of signal amplitude square and combine them together to form the feature vector of each EEG. Because the wavelet package entropy is far less than the sum of absolute value and the sum of signal amplitude square, in order to make sure that the entropy reacts in the third step, we use one thousandth of the sum of absolute wavelet packet coefficients, one hundredth of the sum of signal amplitude square and the wavelet package entropy as the weighted feature vector. Finally, we succeed in distinguishing EEGs between epilepsy and normal by using the optimized AdaBoost whose input is the weighted feature vector. The result shows that the presented method has a high recognition rate, it can identify 96.11% epilepsy EEGs and 99.51% normal EEGs, thus it provides an effective solution for the correct diagnosis of epilepsy.