Lipson Jan, Bernhardt Jeff, Block Ueyn, Freeman William R, Hofmeister Rudy, Hristakeva Maya, Lenosky Thomas, McNamara Robert, Petrasek Danny, Veltkamp David, Waydo Stephen
C8 MediSensors, Los Gatos, California, USA.
J Diabetes Sci Technol. 2009 Mar 1;3(2):233-41. doi: 10.1177/193229680900300203.
In the development of noninvasive glucose monitoring technology, it is highly desirable to derive a calibration that relies on neither person-dependent calibration information nor supplementary calibration points furnished by an existing invasive measurement technique (universal calibration).
By appropriate experimental design and associated analytical methods, we establish the sufficiency of multiple factors required to permit such a calibration. Factors considered are the discrimination of the measurement technique, stabilization of the experimental apparatus, physics-physiology-based measurement techniques for normalization, the sufficiency of the size of the data set, and appropriate exit criteria to establish the predictive value of the algorithm.
For noninvasive glucose measurements, using Raman spectroscopy, the sufficiency of the scale of data was demonstrated by adding new data into an existing calibration algorithm and requiring that (a) the prediction error should be preserved or improved without significant re-optimization, (b) the complexity of the model for optimum estimation not rise with the addition of subjects, and (c) the estimation for persons whose data were removed entirely from the training set should be no worse than the estimates on the remainder of the population. Using these criteria, we established guidelines empirically for the number of subjects (30) and skin sites (387) for a preliminary universal calibration. We obtained a median absolute relative difference for our entire data set of 30 mg/dl, with 92% of the data in the Clarke A and B ranges.
Because Raman spectroscopy has high discrimination for glucose, a data set of practical dimensions appears to be sufficient for universal calibration. Improvements based on reducing the variance of blood perfusion are expected to reduce the prediction errors substantially, and the inclusion of supplementary calibration points for the wearable device under development will be permissible and beneficial.
在无创血糖监测技术的发展过程中，非常希望得到一种校准方法，它既不依赖于个体相关的校准信息，也不依赖于现有侵入性测量技术提供的补充校准点（通用校准）。
通过适当的实验设计和相关分析方法，我们确定了实现这种校准所需的多个因素的充分性。所考虑的因素包括测量技术的辨别力、实验设备的稳定性、基于物理 - 生理学的归一化测量技术、数据集大小的充分性以及建立算法预测值的适当退出标准。
对于无创血糖测量，使用拉曼光谱法，通过将新数据添加到现有校准算法中，并要求（a）在不进行重大重新优化的情况下，预测误差应保持或改善；（b）最佳估计模型的复杂性不会随着受试者的增加而增加；（c）对于那些数据完全从训练集中移除的人的估计不应比总体其余部分的估计差，从而证明了数据集规模的充分性。使用这些标准，我们凭经验确定了初步通用校准所需的受试者数量（30 名）和皮肤部位数量（387 个）的指导原则。我们整个数据集的中位绝对相对差异为 30 mg/dl，92%的数据在克拉克 A 级和 B 级范围内。
由于拉曼光谱法对葡萄糖具有高辨别力，一个具有实际规模的数据集似乎足以进行通用校准。基于减少血液灌注方差的改进预计将大幅降低预测误差，并且为正在开发的可穿戴设备纳入补充校准点将是可行且有益的。
