Multivariate determination of glucose using NIR spectra of human blood serum

Abstract

Multivariate statistical modeling methods have been applied to near-infrared (NIR) spectral data to discriminate glucose concentrations. Specifically, performance levels are compared for principal component regression (PCR) and partial least-squares regression (PLSR) models based on their standard errors of prediction (SEP). NIR spectra of blood serum from 456 individual hospitalized patients were generated using a NIRSystems 6500 spectrophotometer in 2 nm intervals from 400 to 1098 nm. Only the data between 870 and 1098 nm were used for calibration model development and validation. Performance results for the PLSR model (SEP=29.577 mg/dl) were about the same as that obtained with the PCR model (SEP=28.881 mg/dl). >