近红外无创生化分析中模型稳健性研究

发布时间：2018-02-01 05:31

本文关键词： 近红外光谱无创生化分析血流容积光谱相减法模型稳健性　出处：《中国科学院研究生院（长春光学精密机械与物理研究所）》2012年硕士论文　论文类型：学位论文

【摘要】：血液不断流动于人体循环系统之中,对于维持机体新陈代谢、功能调节,以及内外环境间的平衡等方面都起着十分重要的作用。血液生化成分检测作为健康诊断与疾病监控的有效手段,在临床诊疗领域具有非常重要的意义。然而,临床常规采用的血液检验方法多为有创方法,检测周期较长且需用试剂。近红外无创生化检测技术具有无创伤、无感染、无需试剂、便于实时监测等特点,已成为目前国际上的研究热点。目前近红外无创生化分析技术主要面临的问题包括：血液有效信息微弱；人体组织背景干扰严重；血流容积随心脏搏动不断变化。为消除组织背景及脉搏的干扰,课题组提出血流容积光谱相减方法。通过将不同血流容积下测得的两幅光谱相减,可有效扣除皮肤、肌肉等组织背景的干扰,从而提取血流容积变化量对应于血液成分信息的纯净光谱。本论文围绕近红外无创生化分析中的模型稳健性问题展开了研究。为提高血流容积光谱相减法的数据信噪比,考察了人体容积脉搏信号的噪声干扰问题,分析了噪声的主要来源并提出解决方案；为提高血液生化成分定标模型的预测能力,对比分析了不同数据预处理算法、建模方法及其组合下的定标结果,优化模型,提高稳健性。主要研究内容与取得的成果有： 1)应用课题组自行设计的近红外血液无创生化检测系统,通过临床测量实验,无创采集到不同年龄、不同性别志愿者的食指指端容积脉搏波光谱数据,对比分析不同人群的光谱信号差异； 2)分析总结了人体容积脉搏光谱信号的主要噪声来源并提出多种光谱去噪方法,分别采用移动平均平滑、Savitzky-Galay卷积平滑、经验模态分解、小波变换滤波等算法处理脉搏信号,有效地抑制了数据噪声,提升了光谱性能； 3)结合不同数据预处理算法及建模方法建立血液生化成分定量校正模型,对比分析不同定标模型的预测精度,通过优选算法优化模型提高模型稳健性。经实验验证,BP-ANN模型结合经验模态分解方法对HCT及血红蛋白的预测能力较优,预测相关系数可分别达到0.92和0.87,预测标准差分别为1.66%和8.08g/L-1。本文深入分析了近红外无创生化检测中的模型稳健性问题,研究了血流容积光谱相减方法中信噪比的提升方法并优选了血液生化成分建模算法,为近红外无创生化检测技术的实际应用提供了理论和实验基础。
[Abstract]:Blood flow in the human circulatory system, for the maintenance of the body metabolism, function regulation. As an effective means of health diagnosis and disease monitoring, the detection of blood biochemical components plays a very important role in the field of clinical diagnosis and treatment. However, most of the routine blood testing methods are invasive, and the detection period is longer and reagent is needed. Near-infrared non-invasive biochemical detection technology has no trauma, no infection, no reagent. The advantages of real-time monitoring have become an international research hotspot. At present, NIR non-invasive biochemical analysis technology is mainly faced with the following problems: weak effective blood information; The interference of human tissue background is serious; In order to eliminate the interference of tissue background and pulse, the method of blood flow volume spectral subtraction was put forward. Two spectra were subtracted from different blood flow volumes. It can effectively deduct the interference of skin, muscle and other tissue background, so as to extract the pure spectrum of the volume change of blood flow corresponding to the information of blood composition. In order to improve the signal-to-noise ratio (SNR) of blood flow volume spectral subtraction, the noise interference of human volumetric pulse signal was investigated. The main sources of noise are analyzed and solutions are put forward. In order to improve the prediction ability of the blood biochemical component calibration model, different data preprocessing algorithms, modeling methods and their combination of calibration results, optimization model were compared and analyzed. Improving robustness. The main research contents and achievements are as follows: 1) using the NIR blood noninvasive biochemical detection system designed by our research group, we collected the pulse wave spectrum data of index finger volume of volunteers of different ages and different genders through the clinical measurement experiment. The spectral signals of different populations were compared and analyzed. 2) the main noise sources of human body volumetric pulse spectral signal are analyzed and summarized, and a variety of spectral denoising methods are proposed. The moving average smoothing method is used to smooth Savitzky-Galay convolution. Some algorithms such as empirical mode decomposition and wavelet transform filter are used to process pulse signal, which can effectively suppress the data noise and improve the spectral performance. 3) combined with different data preprocessing algorithms and modeling methods, the quantitative calibration model of blood biochemical components was established, and the prediction accuracy of different calibration models was compared and analyzed. The model robustness is improved by optimizing the model by optimal selection algorithm. The prediction ability of BP-ANN model combined with empirical mode decomposition method for HCT and hemoglobin is proved to be better. The predictive correlation coefficient is 0.92 and 0.87, respectively, and the predicted standard deviation is 1.66% and 8.08 g / L ~ (-1), respectively. In this paper, the model robustness in NIR noninvasive biochemical detection is analyzed in depth, and the enhancement method of SNR in blood volume spectral subtraction method is studied, and the modeling algorithm of blood biochemical composition is selected. It provides a theoretical and experimental basis for the practical application of NIR noninvasive biochemical detection technology.
【学位授予单位】：中国科学院研究生院（长春光学精密机械与物理研究所）
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：R-332

【参考文献】