【作者】 张涛；

【导师】 洪文学；

【作者基本信息】 燕山大学 ， 仪器科学与技术， 2012， 博士

【摘要】 帕金森病是人类常见的神经退行性疾病之一，其病程时间长且发病范围广。由于该病病因尚未完全明确，目前所有针对帕金森病的治疗都是控制病症，而无法从根本上进行治愈。但在发病早期就开始接受合理治疗的患者，绝大多数能够延缓病情的发展，生活基本自理。因此，帕金森病的早期诊断不论对于家庭还是社会均具有重大意义。语言障碍是帕金森病的早期症状之一，基于语言障碍的帕金森病诊断是近年来的帕金森病诊断研究热点之一。本文针对帕金森病的语言障碍特征，本文提出利用多维筛组合分类器进行帕金森病的可视化诊断。在解决帕金森病诊断实际问题的同时，完善基于高维数据列向量图表示的可视化组合分类器理论和方法，无论对信息融合和模式识别学术研究还是对基于语音特征的帕金森疾病诊断应用研究都具有重要意义。首先，基于可视化分类器的框架结构，提出多维筛组合分类器框架，并从数据表示、类域生成与权重计算三个方面进行了完善。在数据表示阶段，提出基于色度学混色原理的彩色多元图着色表示，完善了模式识别应用下的多元图表示中类别信息的表示方法；在类域生成阶段，以训练样本的空间单点表示为出发点，提出了基于计算几何的区域主动生长的类界面求取方法。通过对基点区域的主动生长，使得整个表示空间任意区域均可进行类别表示，从而完成分类界面的计算过程；在权重计算阶段，从类空间类别分布特性出发，分别从不同的视觉角度提出基于类空间模糊度与规整度的权重分配方法，分别从统计和结构角度对类空间进行可视化权重计算。对通用数据的综合数据实验表明，多维筛分类器不但具有良好的可视化特性，而且分类性能已经达到或超过主流分类器的水平。其次，研究了帕金森病语音特征与帕金森病临床表现之间的关系。通过实验验证，进一步阐明各语音特征的物理意义及其应用特点，为基于语言障碍检测的帕金森病早期诊断提供了可靠依据。在证明语音特征用于帕金森病诊断的有效性同时，分析了不同特征下不同元音的类间分离度，为基于语音障碍的帕金森病自动诊断奠定了基础。最后，在对基于语言障碍的帕金森病的分类实验中，分别利用帕金森数据集与远程帕金森数据集进行了帕金森病的可视化诊断与病程的判断。实验表明，利用可视化模式识别理论，不但完成了诊断过程的全程可视化，有助于新的诊断指标的发现，而且获得了比经典分类器更高的诊断精度。更多还原

【Abstract】 Parkinson’s disease (PD) is one of highest incident neurodegenerative diseases with along time course and a significant prevalence. As the etiology is currently unkown, thetreatement for this disease is to alleviate but can not to be cured. If the patients could betreated appropriately, most of them could control the disease’s development and preservethe ability of basic self care. As a result of that, early diagnosis of Parkinson’s disease is ofgreat significance both for family and for society. The dysphonia is a typical symptom forearly Parkinson’s disease patient, so the Parkinson’s disease diagnosis in dysphoniameasurement is one of active fields.In this paper, we propose the visual diagnostic method for Parkinson’s disease basedon speech features by visual combined classifer named multi-dimensional filter. Thepurpose of that is not only to explore the way about PD’s diagnose, but also to improve thetheory and method in visual pattern recongnition about multiple graph representation forhigh-dimensional data. It is of great significance both for information fusion, patternrecognition research and Parkinson’s disease diagnosis based on speech features.At first, we proposed the multi-dimensional filter classifier framework based on thetheory of visual classification, and improve it from the data representation, domainconstruction and weight calculation. Stage of data representation, a new type ofrepresentation named chromatic graphical representation is proposed to bridge the gapbetween graphical representation and category distribution. It inherits the merits oftraditional and represent the category information by chromatic of the current sample,which makes it more suitable than the traditional in visual pattern recognition applications;One of the basic issues in pattern recognition is to calculate the boundary betweendifferent categories. Stage of domain construction, a novel method for that based oncomputational geometry named active expansion is proposed. The whole space couldexpress the category information and the boundary is obtained by active expanding forbase points to non-base points; Stage of weight calculation, fuzzy and regularity areproposed for weight based on category spatial distribution characteristics. Two methods analyze the space visual information from the statistical and structure respectively.Comprehensive data experiments show that the multi-dimensional filter classifier not onlyinherits the characteristics about visualization features, but also gets equivalentperformance to the popular classifiers, and outweighs in some dataset.In subsequential, the relationship between speech features and PD symptom isresearched. The physical explainations of speech features and application characteristicsare clarified by experiments, which provide a reliable basis for early diagnosis ofParkinson’s disease base on dysphonia. At the same time, we analyze the differentcharacteristics of different vowel separation between classes, which is the foundation forautomatic diagnosis of Parkinson’s disease.At last, in the experiments about diagnosis the PD based on dysphonia, theParkinson dataset and Parkinson telemonitering dataset are employeed for visual diagnosisand determine the course of the disease. Experiments show that the application of visualpattern recognition, not only completed the entire diagnostic process of visualization toassist the discovery of new diagnostic indicators, also to obtain a higher classificationaccuracy than the classical machine diagnoses.更多还原