These strategies differ in the way to construct the calibration and validation matrices and they had to be carried out to suppress the bias on the quantification of the constituents in the polymer blend.
For example, three randomly generated calibration and validation matrices were combined with the Raman spectral data that had been transformed using a first derivative and EMSC.
The NIR spectral data required four to six factors to explain the variance and had between one and five samples characterized as an outlier, depending on the calibration and validation matrices.
Additionally, the metrics listed in Tables 2 and 3 are the average results obtained from using three randomized calibration and validation matrices for each spectral transformation (Additional file 1: Table S3).
Of the 195 calibration samples, only two were characterized as outliers after thoroughly evaluating the leverage, Hotelling T, residual variance, and X-Y distribution statistical plots, although this characterization was dependent on the randomized calibration and validation matrices.
This tactic permitted the development of completely independent predictive models, whereas the use of one set of calibration and validation matrices for all spectroscopic data sets may have introduced a level of undesirable bias into the models.
The calibration R 2 values ranged from 0.83 ± 0.01 to 0.845 ± 0.003, four to six factors were required to explain the variance, and between two and five samples were shown to be outliers, dependent on the respective calibration and validation matrices (Table 2, Additional file 1: Table S3).
All data matrices used for training, testing and validation are available in Additional file 2: Table S2.
Let us define g and T as ξ′ matrix and a given validation tumor respectively or a given training matrix and a corresponding given training tumor respectively.
The use of the matrix distance to compare various AIP has many advantages for methodology development and validation.
