Improvements to the Prediction Accuracy of a Noninvasive Mid-Infrared Glucose Sensor

Abstract

A glucose sensor implemented using pulsed mid-infrared quantum cascade laser technology constitutes a promising approach to sensing glucose levels in patients, and we have demonstrated a system that is able to correctly gauge glucose levels over 70% of the time, across multiple users, using a system that sweeps monochromatic mid-infrared light over a range of wavenumbers, then analyzes the spectra using machine learning techniques to produce a glucose concentration reading. To improve this system, inefficiencies in both the optomechanical equipment and computational data-processing program needed to be addressed. We first show that a smaller integrating sphere and a larger-area detector both theoretically increase the detector signal, all else being equal, and present the results of installing a smaller integrating sphere into our system.We also describe our implementation of a stepped-wavenumber scan, a data collection method that enables more accurate measurements with a richer set of metadata, to replace the swept laser scan procedure, and analyze how various classes of noise visible in our system influence the optimal averaging time constants of the system. These changes to the optical components and the data collection strategies allow us to collect data with a stronger signal, and also provide us with additional information about the extent of disturbances acting on the system. The improvements we describe are important to the machine learning-based data postprocessing aspect of our system, and will allow it to better handle the relatively small in vivo training sets that the system is likely constrained to use.