Photonic Monte Carlo Analysis and Deep Learning Predicting the Performance of Non-Invasive Glucose Detection Using Compact NIR Spectrometers

Continuous glucose monitoring can be very important in preventing life-threatening dehydration and diabetic comas. For this purpose, the analysis device has to be compact and easy to use. In this paper, neural networks and optical Monte Carlo (MC) analysis are used to study the feasibility of predicting the glucose concentration non-invasively through the skin using compact near infrared spectrometers. MC is used to simulate light diffuse reflectance returning from the skin under wideband light illumination, in the spectral range of 900 nm to 2400 nm, and a 3-layer model is used to describe the skin. The presence of glucose in the skin layers is added into the skin model and a dataset of 6100 spectra is generated varying glucose and water concentrations. The dataset is then modified by superimposing artifacts mimicking the performance of the compact spectrometer such as a signal-to-noise ratio from 200 to 800, a wavelength accuracy of 2 nm and a spectral resolution from 5 nm to 60 nm. A training subset of the data is fed to a neural network and the network is trained and tuned to predict glucose concentrations at different levels of device performance. The RMSE of prediction varies between 3.1 and 14.6 mg/dL, while the resulting R
²
score is in the order of 0.98-0.99 demonstrating the potential of this setup.