Accurate Post-Calibration Predictions for Noninvasive Glucose Measurements in People Using Confocal Raman Spectroscopy

Abstract

In diabetes prevention and care, invasiveness of glucose measurement impedes efficient therapy and hampers the identification of people at risk. Lack of calibration stability in non-invasive technology has confined the field to short-term proof of principle. Addressing this challenge, we demonstrate the first practical use of a Raman-based and portable non-invasive glucose monitoring device used for at least 15 days following calibration. In a home-based clinical study involving 160 subjects with diabetes, the largest of its kind to our knowledge, we find that the measurement accuracy is insensitive to age, sex, and skin color. A subset of subjects with type 2 diabetes highlights promising real-life results with 99.8% of measurements within A + B zones in the consensus error grid and a mean absolute relative difference of 14.3%. By overcoming the problem of calibration stability, we remove the lingering uncertainty about the practical use of non-invasive glucose monitoring, boding a new, non-invasive era in diabetes monitoring.

Keywords: calibration stability; diabetes; in vivo Raman spectroscopy; multivariate data analysis; non-invasive glucose monitoring; portable sensor; tissue diagnostics.

Figure 1

Non-invasive glucose sensor. (a) Novel, production-ready, portable, stand-alone, and Raman-based device configured for NIGM. (b) Schematic optical layout. (c) Examples of recorded thenar spectra from five subjects with different skin colors, according to the Fitzpatrick scale, where type I and type V correspond to the lightest and the darkest skin complexions, respectively. Spectra are vertically offset for clarity.

Figure 2

Calibration stability. Comparison between the daily mean of the measured and reference glucose value and the subject-wise, average RMSE for the 160 subjects for a validation period of 15 days. The bars on the RMSE curve represent the standard deviation.

Figure 3

Measured glucose concentrations plotted as a function of reference values in a consensus error grid for all type 1 (a) and type 2 subjects (b). The reference glucose value is obtained as the average of two blood glucose measurements (Contour Next One, Ascensia), whereas the corresponding glucose measurement is the result of the PLS regression model applied to three pre-processed NIGM spectra.

Figure 4

Histograms of subject-wise RMSE values for (a) 137 subjects with type 1 diabetes and (b) 23 subjects with type 2 diabetes.

Figure 5

Regression vectors from the individual prediction models. The top shows the Raman spectrum for glucose and the average regression vector obtained from the PLS prediction models. The regression vector is seen to mimic the significant peaks in the glucose spectrum. This is consistent for all 160 subjects as demonstrated on the color-coded map.