Data preprocessing and partial least squares regression analysis for reagentless determination of hemoglobin concentrations using conventional and total transmission spectroscopy.

Visible-near infrared spectroscopy was successfully used for the determination of total hemoglobin concentration in whole blood. Absorption spectra of whole blood samples, whose hemoglobin concentrations ranged between 6.6 and 17.2 g/dL, were measured from 500 to 800 nm. Two different types of transmission were measured: conventional transmission spectroscopy which collected primarily collimated radiation transmitted through the sample, and total transmission spectroscopy which used an integrating sphere to collect all scattered light as well. Different preprocessing techniques in conjunction with a partial least squares regression calibration model to predict hemoglobin concentrations were applied to the above two types of transmission. Depending on different preprocessing methods, the standard error of predictions ranged from 0.37 to 2.67 g/dL. Mean centering gave the most accurate prediction in our particular data set. Preprocessing methods designed for compensation of the scattering effect produced the worst results contrary to expectations. For univariate analysis, better prediction was achieved by total transmission measurement than by conventional transmission measurement. No significant difference was observed for multivariate analysis on the other hand. Careful selection of the data preprocessing methods and of the multivariate statistical model is required for reagentless determination of hemoglobin concentration in whole blood.