FT-MIR modelling enhancement for the quantitative determination of haemoglobin in human blood by combined optimization of grid-search LSSVR algorithm with different pre-processing modes

Abstract

Haemoglobin (HGB) is an important factor in determining anaemia and iron nutrition for human health. The quantitative determination of HGB in human blood is performed by the rapid analytical tool of Fourier transform mid-infrared (FT-MIR) spectrometry with its chemometric algorithms. Least-squares support vector regression (LSSVR) is utilized for nonlinear modelling. For the enhancement of modelling, we propose that the grid-search technique should be applied to tune the parameters of LSSVR modelling. Moreover, we constructed a framework for discussing the separate and combined use of the spectral pre-processing methods of multiplicative scatter correction (MSC), standard normal variate (SNV) and Savitzky–Golay smoother (SGS), in which the SGS parameters were set to be tunable in a certain designated range. The performances of different pre-processing modes were evaluated in combination with grid-search LSSVR modelling. To obtain stable results, grid-search LSSVR models and pre-processing modes were established based on the average predictive results of 30 different calibration-validation divisions. These analytical methods were carried out in the FT-MIR fingerprint region of human blood HGB and compared with those carried out in the full-scan region. Results show that the optimized model appears in the fingerprint region. In the evaluation of test samples, the designated optimal model exhibits a root mean square error of testing (RMSET) of not more than 6% of the mean chemical value with a correlation coefficient higher than 0.9. This study shows that the combined optimization of a grid-search LSSVR algorithm with different pre-processing modes has the potential of improving the predictive abilities of FT-MIR spectroscopic analysis of HGB in human blood.