近红外漫反射光谱具有无创伤、 连续、 无感染、 速度快等诸多优势, 在人体成分无创伤检测方面有很好的应用前景。 但是在测量过程中, 随机噪声、 干扰组分以及检测条件的改变等容易导致异常光谱。 判定并剔除异常光谱对于提高近红外无创血液成分检测的可靠性具有重要意义。 首先分析了近红外漫反射光谱无创血糖检测中可能出现的异常数据类型, 提出了一种综合利用马氏距离、 光谱残差和化学值残差三个指标构造三维空间对样本集进行检验的三维坐标异常数据判定方法。 其次, 针对三层皮肤组织模型, 在参数中设置人为失误、 极端成分含量以及异常温度变化的样本, 通过蒙特卡罗（MC)模拟程序得到一组正常模拟数据以及一组包含化学值异常和光谱异常的模拟数据, 并利用三维坐标法进行异常数据的判定。 结果显示, 该方法能识别出全部异常样本, 剔除这些异常样本后, 偏最小二乘（PLS）校正模型的交互验证均方根误差（RMSECV）由212 mmol·L-1降低到11 mmol·L-1, 初步验证了该方法的可行性。 进一步, 对三位受试者开展了口服葡萄糖耐量试验（OGTT）, 通过在测量受试者血糖参考值的同时同步采集其手指部位的漫反射光谱, 获得了三组在体实验数据。 并利用三维坐标法和蒙特卡罗交互验证法进行异常数据的判定和剔除, 最后建立PLS模型比较两种异常数据判别方法的效果: 剔除三维坐标法识别出的异常数据后, 三组样本建立的校正模型的决定系数显著提升, RMSECV平均值由21 mmol·L-1降低至08 mmol·L-1, 效果优于蒙特卡罗交互验证法的结果。 这些结果表明, 基于马氏距离、 光谱残差和化学值残差的三维坐标异常数据判定方法能有效识别近红外无创血糖测量中的异常数据, 在在体成分检测应用中有显著优势。

Near-infrared diffuse reflectance spectroscopy has many advantages, such as being non-invasive, continuous, non-infectious, fast, in the non-invasive detection of body components. It has a great prospect in the application of blood glucose measurement in vivo. However, outliers often occur in the process of measurement due to the random noise, the change in interference components or the measurement conditions. Therefore, it is of great significance to eliminate the outliers in the near-infrared spectroscopy and thus improve the reliability of non-invasive blood components measurement. In this paper, the types of outliers that may occur in the blood glucose sensing by near-infrared diffuse reflectance were analyzed, and a three-dimensional coordinate outlier determination method based on the three-dimensional space constructed by the residual of chemical value, the Mahalanobis distance and the spectral residuals was proposed firstly. Then, it was used to discriminate the outliers in the simulated spectra of three-layer skin model by Monte Carlo program, where the abnormal data was obtained by adding the artificial errors, abnormal chemical values and abnormal temperature changes in the parameters setting in Monte Carlo simulation. All the outliers could be found successfully by the three-dimensional coordinate outlier determination method, and the root-mean-square error of cross-validation (RMSECV) of the Partial Least Square (PLS) model was reduced from 212 to 11 mmol·L-1 after the removal of outliers. Further, the oral glucose tolerance tests (OGTTs) of three volunteers were carried out, where three groups of experimental data were obtained by measuring the reference blood glucose concentrations and collecting the diffuse reflectance of finger synchronously, and Monte Carlo Cross-Validation outlier detection method and three-dimensional coordinate method were used to detect the outliers, respectively. Results showed that, after the removal of outlier by the three-dimensional coordinate method, the coefficient of determination of calibration model increased significantly, and the average RMSECV value of calibration model for three sets of samples was reduced from 21 to 08 mmol·L-1, which was better than that of MCCV method. All these results indicated that, three-dimensional coordinate method can effectively determine the outlier in the near-infrared diffuse reflectance and it’s more suitable for the non-invasive blood glucose measurement in vivo by near-infrared diffuse reflectance spectroscopy.