Analyzing Multivariate Calibration Techniques for Glucose Level Prediction in Non-invasive Human Tongue Spectra

Non-Invasive Glucose Monitoring devices use Multivariate Calibration (MC) methods to estimate glucose concentration in blood. The accuracy of methods depends on spectral data obtained from tongue-to-spectrometer interface. In this work we examine four widely used MC methods, they are: classical least square (CLS), inverse least square (ILS), principal component (PC) and partial least square (PLS). We discuss the impact of factor selection on the prediction of response for first overtone transmission spectra collected across human tongues. Results are exposed for various signal-to-noise ratio (SNR) and different factor values. We explicitly tackle the issue of low SNR in tongue-to-spectrometer interface. We show that CLS outperforms factor based regression techniques where SNR is as low as 30 dB. Our results are useful in calibration models that are used to predict vivo glycemia from human tongue spectra.