Figure 5. The Cross-Validation Percent Residual for Moisture.
The calibration curve and percent residual for moisture are shown in Figures 2 and 3. The correlation coefficient (how the
data is fit by a straight line; 1.0 is theoretically perfect) is 0.998 with a Root Mean Square Error of Calibration (RMSEC)
of 0.005. Both parameters describe a good fit of the spectral data to the reference data. Cross-validation is an excellent
way to gauge the stability of a certain method. One by one, each calibration standard is removed from a calibration and predicted
against the remaining nine standards. If the method is robust, the predicted values of these removed standards will not be
very far from their original values. If the calibration seems to get significantly worse upon cross-validation, then the method
should be reviewed and perhaps have more standards added. Cross-validation for moisture in thrombin gave a correlation coefficient
of 0.984 with a Root Mean Square Error of Cross Validation (RMSECV) of 0.018. Although cross-validation indicated some change
in the line fit from the calibration curve, the RMSECV was only a few times greater than the RMSEC, suggesting a relatively
stable method. The cross-validation curve for moisture is shown in Figure 4.
Figure 6. The Calibration Curve For Potency.
Potency analysis was conducted by correlating primary potency data derived from titration scattering measurements with spectral
data from a second set of ten thrombin samples. These samples had potency values from approximately 29,000 to 33,000 units.
The pretreatment for this data set was slightly different than that for moisture. Here, a Partial Least Squares (PLS) algorithm
was used on the second derivative spectra with a Norris 9,5 smoothing filter. The pathlength compensation used was a multiplicative
scatter correction. The analysis region for potency was 6,000 cm-1 to 6,800cm-1 , which is in between, but not part of, the water resonances. The calibration curve for this case (Figure 6) was excellent,
with a correlation coefficient of 0.999 and an RMSEC of 21.9. The percent residual (Figure 7) for nine of the calibration
standards showed prediction values for potency of ±0.09%, although sample 18582 had a residual value of –0.18. The Predicted
Error Sum of Squares (PRESS) also showed a reasonable trend (Figure 8). For a reasonable calibration curve, the PRESS plot
should start high, tend towards a minimum and then stay flat or increase slightly. Cross-validation for this calibration showed
the residual increase to approximately ±2.0%. However, the high and low standards had a greater degree of error because they
were under-represented in the calibration.