Improving Accuracy of Non-Invasive Glucose Monitoring Through Non-local Data Denoising

Correlation and clinical interpretation, in respect to the true glucose value of patient is imperative for optimum therapy and disease management. Accuracy of optical glucometer is hampered by many debilitating factors such as concentration range, sampling environment, tongue-to-spectrometer interface, changes in wavelength, polarization or intensity of light, to name a few. Regression techniques are used in such devices to build patient glucose model. This work is an extension to our previous work regarding multivariate calibration for glucose level prediction in noninvasive human tongue spectra. Here, we present our results for noise reduction and data conditioning during glucose spectrum isolation phase. We embed our 'Indicator Function (IF)' scheme into two popular techniques known as Outlier Sample Removal (OSR) and Descriptor Selection (DS). Methodology is tested on dataset 'OCATNE20' obtained from a public domain website and results are compared at both OSR and DS for a wide range of blood serum samples. Our results show that outlier samples identification and removal in early stage significantly increase the prediction of unknown samples typically in the range of 7.95% to 9.84%.