Near Infrared Multispectral Imaging for Non-invasive Continuous Glucose Monitoring: a feasibility study

According to the International Diabetes Federation, the worldwide diabetes prevalence was 10,5% (537 million people) in 2021. The Near Infrared (NIR) Spectroscopy is raised to be a promising method for painless, reliable non-invasive glucose continuous monitoring. The aim of this work is the characterization of a multispectral image sensor to predict non-invasively the concentration of biological substances. The photosensor exploited is the STMicroelectronics VD56G3 (aka FOXY), enhanced with a filters array (64 total wavelengths) in the NIR range 780-1062 nm. First, the cross-channel interference within the NIR filters array is observed. This optical phenomenon leads to the distortion of the prototype spectral sensitivity and negatively affects the final prediction. A theoretical method to remove this phenomenon is discussed. Due to instrumentation limits, the cross-channel interference is not removed, but maximally reduced. The optimal configuration is proved to be a light beam exactly perpendicular to the photosensor surface. A Convolutional Neural Network (CNN) classifies the angle of the incident light beam to verify the correct configuration. With an accuracy of 90 %, the CNN Classifier proves to be reliable. The Glucose and the Alcohol are the two target solutes. The glucose dataset covers aqueous concentrations from 44 to 500 mg/dl, to overestimate the human tolerable glycemia level. The alcohol aqueous solutions range 0.03-0.1% g/L, to explore the legal Blood Alcohol Level (BAC) interval. The relationship between the concentration and NIR spectra is investigated. The Partial Least Square (PLS) regression is used for the linear regression. The Extra Trees (ET) and the Multilayer Perceptron (MLP) regressors explore the non-linear method. The Genetic Algorithm (GA) is used for the Feature Selection (FS). How using the major contributing features selected by the GA can reduce cross-channel interferences and simplify the model to embed the final system is discussed. As expected, a linear regression model is inappropriate for the Glucose concentrations: it produces negative values for R2. Nonlinear ET and MLP regressors enhance the prediction during the learning process. The MLP Regressor reaches R2 of 0.76±3e-5 (Mean Absolute Percentage Error (MAPE): 25±0.2 %) in Glucose Test Set, with all mispredictions outside of the physiological glycemia range (> 200 mg/dl). The performances drop down when completely unseen data are used (Validation Set: negative R2 and MAPE: 94±30%). The GA selects 34 (out of 64) major contributing features. This subset of feature does not improve the MAPE and the R2 of the models, but significatively decreases the standard errors in predictions. The alcohol concentration is highly predictable by wavelengths in the NIR range 780-1062 nm, even with a linear regression method. The PLS reaches R2: 0.87±0.12 (MAPE: 9.4±0.14 %) in Validation Set. The FS selects a subset of 29 features without a loss in performances (R2: 0.86±0.13, MAPE: 9.6±0.1%). The nonlinear ET regression method achieves better performances (R2: 0.84±0.13, MAPE: 5.6±0.07 %) in the Validation set. After the FS, R2: 0.74±0.13, MAPE: 7±0.06%. The model proposed is highly performant in the non-invasive continuous monitoring of substances with a sensible absorbance in the NIR range 780-1062 nm (Alcohol). This NIR range does not include any appropriate feature able to reliably fit the nonlinear relationship between the glucose concentration and the spectra.