Temperature insensitive prediction of glucose concentration in turbid medium using multivariable calibration based on external parameter orthogonalization

The measurement accuracy of non-invasive blood glucose concentration (BGC) sensing with near-infrared spectroscopy is easily affected by the temperature variation in tissue because it would induce an unacceptable spectrum variation and the consequent prediction deviation. We use a multivariable correction method based on external parameter orthogonalization (EPO) to calibrate the spectral data recorded at different temperature values to reduce the spectral variation. The tested medium is a kind of tissue phantom, the Intralipid aqueous solution. The calibration uses a projection matrix to get the orthogonal spectral space to the variable of external parameter, i.e. temperature, and then the useful spectral information relative to glucose concentration has been reserved. Even more, training the projection matrix can be separated to building the calibration matrix for the prediction of glucose concentration as it only uses the representative samples’ data with temperature variation. The method presents a lower complexity than modeling a robust prediction matrix, which can be built from comprehensive spectral data involved the all variables both of BGC and temperature. In our test, the calibrated spectra with the same glucose concentration but different temperature values show a significantly improved repeatability. And then the glucose concentration prediction results show a lower root mean squared error of prediction (RMSEP) than that using the robust calibration model, which has considered the two variables. We also discuss the rationality of the representative samples chosen by EPO. This research may be referenced to the temperature calibration for in vivo BGC sensing.