Automatic Detection Of Interictal Epileptiform Discharges Based On Perceptual Organizational Principle And Support Vector Machine

Electroencephalogram (EEG) examination is one of the most effective methods for epilepsy diagnosis. The major task of EEG examination is analizing the abnormal activity of brain including spike wave, sharp wave, spike and slow wave complex, sharp and slow wave complex, spike rhythm. Usually, EEG signals of suspected epilepsy patients are recorded at least24hours, and then doctors should observe the recordings page by page, describe the EEG activities, sleep and abnormal wave to conclude whether the signals are abnormal or not. It is a very tedious and time consuming work for doctors. Presently, it is still irreplaceable for the manual interpretation of Interictal Epileptiform Discharges (IED) in the abnormal waves. Therefore, studying of visual cognitive principles and extracting the feature via computing are the main direction of automatic IED detection. It has important application and theory values.In this paper, I summarized the automatic IED detection methods comprehensively and systematically, and made an intensive study of statistical learning theory and fast algorithm of support vector machine for large scale training samples. This thory lays the foundation of pattern recognization, and supply the important guiding ideology of the approaches in this paper. Support vector machine is a quite efficient method of pattern recognition. Lastly, recognition problem is changed to a constrained quadratic programming problem. Sequential minimal optimization algorithm is a rapid method to solve the problem of large-scale samples classification. This paper proposes three new SMO algorithms, and the expriments show that the performance can be largely improved. For automatic IED detection, the perceptual organization principle was applied to analyze the EEG signals by using time-series sequence merging method. It improved the processing speed and accuracy of automatic detection system. The values of sensitivity and specificity are higher than any of current published papers.The main topics studied in this thesis are as follows:Proposed an improved SMO method based on dynamic filtration strategy (DFSMO). It cosists of two processes:roughing and cleaning. The experiments indicate that the speed is increased by70%for UCI Adult datasets and increased by35%for Web datasets.Proposed an improved SMO method based on transcendent knowledge (TKSMO). It adopts a new stop condition for I0rearching. The experiments indicate that the speed is greatly increased and the generalization is more stable than DFSMO algorithm.Proposed a new SMO method to determine the penalty coefficients for different samples. Simulation results from applying the proposed method to binary classification problems show that the generalization performance of the proposed method was better than standard SVM algorithm in the cases that the sizes of binary sample sets1) were selected in proportion;2) were the same;3) were quite different.Proposed a new automatic detection method of IED based on the merger of the increasing and decreasing sequences (MIDS) that aims to improve the detection rate of IED. Firstly, the definitions of increasing and decreasing sequence, as well as complete and incomplete waves, are reviewed to highlight characteristics of clinical visual detection. The merger rules and algorithm are consequently proposed for EEG signals processing in time-domain. Experimental results demonstrate that MIDS detection method for rhythm wave and slow wave are very close to the visual measurements results. Then MIDS detection method is performed for IED fragments based on features in time-domain. The results show that most IED fragments are recognized, although with some false detection of non-IED fragments. To reduce such false detection rate, Support Vector Machine (SVM) was applied with17characteristics and trained with232fragments from3patients’ EEG recordings. Clinical EEG recordings of32suspected epilepsy patients were analyzed and95.9%of the IED fragments marked by clinicians were successfully detected. The results show that the proposed algorithm performs well in IED detection and is a promising candidate in assisting clinicians’ epilepsy diagnosis.On the basis of the merging of EEG signals, a new IED detection method is proposed. The first step of this new method is to establish a database by selecting peaked wave fragments. Then, the similarity between pending test fragment and peaked wave samples in the database is calculated. When the maximum similarity is greater than a certain threshold, the fragment is judged to be a peaked wave. In this research,92IED fragments from4suspected epilepsy patients are collected to establish the sample database. The proposed method was tested on EEG recordings from other31suspected patients. The results show that98.36%of the IED fragments marked by doctors were detected. The experimental results show that this method performs well at IED detection in the clinical EEG data. The similarity is measured based on the comparison between fragments of different time length and can be viewed as a novel approach for the detection of typical EEG waveform.FFT, wavelet analysis, principal component analysis, ndependent component analysis and hilbert-huang translate are not used in this paper. The perceptual organization principle was concretized as time-series sequence merging method, compared with advanced signal processing algorithms, it has outstanding features as follows:(1) it can analyze non-linear and non-stationary signals;(2) it is completely self-adaptive to analyze waves of different rhythm and different amplitude;(3) it is not restricted by Heisenberg uncertainty principle and suitable for analysis of mutations in the signal;(4) it is suitable for slow wave analysis, it is more difficult to detect slow wave than peaked wave because its shape, amplitude and frequency have wider variation range;(5) small computation amount is especially be appropriate for long EEG recording;(6) it can calculate the feature values of amplitude and frequency of single wave, the merger wave has clear physical meanings. Currently, a large amount of diagnosis experiences are obtained from vision, the proposed method can easily apply the visual experiences to the analysis of EEGs. A unified discrimination rule is proposed in this paper for the detection of four kinds of IED waves. The results show that our method has low missing detection rate and high stability. This method can be applied to detect other waves, such as alpha, beta, theta, gammar, spindle and slow waves.