Measurement of glucose and other analytes in undiluted human serum with near-infrared transmission spectroscopy

Abstract

Near-infrared calibration models are described for the measurement of total protein, albumin protein, globulin protein, triglycerides, cholesterol, urea, glucose, and lactate. Spectra are collected in triplicate over the 5000–4000 cm −1 spectral range with a 2.5 mm optical path length for 242 undiluted human serum samples. Calibration models are generated for each analyte by performing partial least-squares (PLS) regression on raw and digitally filtered spectra. Models are optimized individually for each analyte by considering spectral range, number of model factors and width/position of a Gaussian shaped filter response function for a digital Fourier filter. Accurate measurements are demonstrated for each analyte except lactate which is below the detection limit under our experimental conditions. Standard error of prediction (SEP) for glucose is 23.3 mg/dl (1.29 mM). Relevance and stability of the glucose calibration model are examined by evaluating the accuracy of glucose predictions from 50 human serum samples collected on a modified spectrometer nineteen months after the calibration data were collected. In addition, these 50 subsequent samples were treated in a blind manner by withholding their glucose values until all predictions were complete. Results indicate a slight positive bias in the predictions corresponding to a minor instability in the model. This instability is likely due to changes in the spectrometer hardware. Nevertheless, the strong correlation between predicted and actual glucose levels in these blind samples strongly suggests that this calibration model is based on information particular to glucose.