Comparison of Different Approaches of Calibration Transfer: Application to Portable Spectrometers for Measuring Blood Hematocrit of Cattle

Even a well-performing calibration model built on an instrument is normally difficult to exhibit the best performance when it is installed in other instruments due to the unavoidable individual difference of instruments. Since building the own calibration models for each instrument is practically impossible, a proper method for calibration transfer is desired. We weighed methods for building a robust model and for achieving calibration transfer among one master and two slave instruments designed for measuring hematocrit (Hct) of blood of cattle by round robin tests with determination of optimum latent variables (LVs), preprocessing methods, wavelength band selection, skew and bias correction (SBC), and piecewise direct standardization (PDS).
The total 581 cattle blood samples were obtained from 365 cattle grazed in pastures in the east Japan and from five cows raised in the institute, whose Hct values decreased over time by the artificial infection with hemoprotozoan parasite (Theileria orientalis). Additionally, artificially diluted or concentrated samples from blood of a healthy cow were prepared in 20 different levels of Hct values evenly distributed from low to high. The latter group of samples had been oxidized and thus perfectly stable for PDS modeling. All of the blood samples in test tubes were measured by ‘Master’, ‘Slave A’, and ‘Slave B’ sequentially in the short wavelength NIR range of 700-1050 nm with a 1-nm interval. The instruments were originally designed for field measurements based upon a multichannel spectrometer (S-2930, SOMA Optics, Ltd.,) equipped with a touch panel and Li-ion battery. The popular method of centrifugation was applied for measuring the Hct values. All computations were performed in Matlab (Mathworks) with PLS_toolbox (Eigenvector Research, Inc.). Different preprocessing methods, including: 1st or 2nd order derivative, SNV and combinations of them, were applied to the raw spectra. The Hct values were converted into a logarithmic scale to eliminate the nonlinearity owing to the light scattering of blood cells.
By use of the selected range of 850-1050 nm, the accuracy and the transferability were dramatically improved with many preprocessing methods. This result implies the famous absorption band of deoxidized hemoglobin peaked at 760 nm is a destabilizing factor for the transferability. Both SBC and PDS gave their better performance, which specifically means both high accuracy and high transferability, when models were preprocessed by 1st derivative or nothing. However, considering the complicated preparation of perfectly stable blood samples, SBC and PDS are not necessary for our case. Our results show a fortunate case in which calibration transfer can be achieved even better by the effort on making model robust than resorting to the sophisticated methods of standardization, that is a good example of calibration transfer for instruments having relatively small differences (e.g., small wavelength shift) and for predicting a large change in absorbance (i.e. Hct).