Optimizations of preprocessing and wavelength selection in predicting human total hemoglobin concentrations based on VIS/NIR spectroscopy

The importance and effects of data preprocessing and wavelength selection were investigated in predicting total hemoglobin concentrations from absorption spectra. Spectra of the 1nm interval between 500∼900nm were measured from the whole blood samples taken from 165 patients whose hemoglobin concentrations ranged between 7∼17 g/dl. The concentrations were predicted using the partial least squares regression. A total of 18 different combinations of preprocessing were tested. The partial least squares regression analysis provided quite different results depending on preprocessing methods and a wide range of prediction accuracy was obtained. For example, the sum of squares of difference ranged from 6∼18.6, R² varied from 0.8333 to 0.9477 and the root mean squared errors were from 0.5504∼0.996 g/dl. The best result was obtained from the data processed by Linear Regression Baseline Fitting, Unit Area Correction, Mean Centering and Variance Sealing. Instead of using all wavelengths in the broad-band spectra, a discrete number of wavelengths were selected to predict the concentrations using our algorithm, which will be advantageous in developing compact and less expensive commercial devices. It proves that a careful selection of wavelengths can provide a comparable accuracy obtained from using the broad-band spectra. For our particular experimental data, the measurement from only three discrete wavelengths could provide excellent results.