Development of a portable system for non-invasive glucose sensing in the mid-infrared

Abstract

The development of a portable non-invasive glucose monitoring system would benefit the lives of millions of diabetics, who currently rely on a painful finger prick to monitor glucose levels. In particular, we focus on using mid-infrared light to detect glucose levels in the interstitial fluid of the dermis layer of skin. Past research has shown that Quantum Cascade lasers can penetrate to the dermis layer and produce in vivo backscattering spectra amenable for glucose absorption. In addition, we have been able to use partial least squares regression to predict in vivo glucose concentrations with 70% clinical accuracy on average in human subjects. Currently, we have improved the experimental setup so that we can take spectra with high granularity and more stability at room temperature in a non-laboratory setting. Here we focus on improving the prediction analysis by optimizing the data processing and determining the best prediction model. From our results, we show the advantages of using discrete wavelet transformations and differentiation to improve the runtime and accuracy of the prediction analysis. Furthermore, we developed prediction models that capture both the linear and nonlinear relationships between the glucose absorption features and glucose concentrations. In particular, we implement neural network partial least squares (NNPLS) to increase the rate of clinical accuracy by 50% and decrease the root mean square error by 60%. Overall, we have shown that the current experimental system is robust enough to predict clinically accurate glucose concentrations in a diverse set of individuals in a nonlaboratory setting. While there is still room for improvement, these results indicate the feasibility of developing a portable non-invasive glucose monitoring system.