12 Method Development
287
chemical content in a sample is varied along with the concentration of other ingredients. There are many other types of designs that can be used (central composite
designs, fractional factorial designs, mixture designs, spiral designs, D, and I optimal
designs) and attempts at comparing their performance have been published [14, 15].
In situations where the samples are expected to vary both in chemical and physical
properties, designs can be augmented through exposure to the conditions (i.e., calibration set is scanned multiple time after exposition to various moisture environments
to build moisture robustness). Nested designs can also be employed where at each
chemical design point (i.e., Fig. 12.4, point (1, 1, −1)) a design for other parameters is built. This approach is very common for pharmaceutical oral solid dosage
forms such as tablets where the model is expected to be robust to tablet thickness
and density variability.
The use of artificial samples is very attractive as it allows the rapid development of
models for a fraction of the cost of running a manufacturing line. However, care must
be taken to ensure the representativity of the samples designed at a scale different to
what the model is expected to encounter during commercial use. In situations where
the difference between small- and large-scale production samples is significant and
cannot be accommodated for by spectral preprocessing, much of the work may not
be relevant and a new approach to model building may need to be designed. Attempts
to bridge the difference in sample matrix between small and large scale for powder
mixing have been published [16].
12.3.5.3 Model Optimization
NIR spectroscopy relies on the use of multivariate regression methods such as PCR,
PLS, CLS, ANN, or SVM to relate the absorbance values with the reference values.
The description of each method and situation of use is beyond the scope of this
chapter, and readers should refer to the chemometrics chapter, but all these methods
will generate a set of coefficients that can be used to relate a new spectrum to its value
in the parameter of interest as outlined in Fig. 12.1. Once a method is selected, much
of the model optimization relies on determining the calibration set membership, the
wavelength or wavenumber range, the pretreatment of the spectra, and setting the
model complexity (i.e., number of factors for PCR and PLS, network structure and
neuron numbers for ANN, kernel type for SVM).
The optimization is an iterative empirical process and should be performed to
meet the method requirements outlined in the ATP (Table 12.1). Below are some
elements to consider:
• Variable range: it should be optimized to ensure specificity and robustness. For
instance, the development of a model for moisture should use spectral ranges at
1450 nm and or 1920 nm corresponding to the -OH absorption band. But the
overlap of chemical absorption bands in the NIR range usually means that large
variable ranges are included in the model.
287
chemical content in a sample is varied along with the concentration of other ingredients. There are many other types of designs that can be used (central composite
designs, fractional factorial designs, mixture designs, spiral designs, D, and I optimal
designs) and attempts at comparing their performance have been published [14, 15].
In situations where the samples are expected to vary both in chemical and physical
properties, designs can be augmented through exposure to the conditions (i.e., calibration set is scanned multiple time after exposition to various moisture environments
to build moisture robustness). Nested designs can also be employed where at each
chemical design point (i.e., Fig. 12.4, point (1, 1, −1)) a design for other parameters is built. This approach is very common for pharmaceutical oral solid dosage
forms such as tablets where the model is expected to be robust to tablet thickness
and density variability.
The use of artificial samples is very attractive as it allows the rapid development of
models for a fraction of the cost of running a manufacturing line. However, care must
be taken to ensure the representativity of the samples designed at a scale different to
what the model is expected to encounter during commercial use. In situations where
the difference between small- and large-scale production samples is significant and
cannot be accommodated for by spectral preprocessing, much of the work may not
be relevant and a new approach to model building may need to be designed. Attempts
to bridge the difference in sample matrix between small and large scale for powder
mixing have been published [16].
12.3.5.3 Model Optimization
NIR spectroscopy relies on the use of multivariate regression methods such as PCR,
PLS, CLS, ANN, or SVM to relate the absorbance values with the reference values.
The description of each method and situation of use is beyond the scope of this
chapter, and readers should refer to the chemometrics chapter, but all these methods
will generate a set of coefficients that can be used to relate a new spectrum to its value
in the parameter of interest as outlined in Fig. 12.1. Once a method is selected, much
of the model optimization relies on determining the calibration set membership, the
wavelength or wavenumber range, the pretreatment of the spectra, and setting the
model complexity (i.e., number of factors for PCR and PLS, network structure and
neuron numbers for ANN, kernel type for SVM).
The optimization is an iterative empirical process and should be performed to
meet the method requirements outlined in the ATP (Table 12.1). Below are some
elements to consider:
• Variable range: it should be optimized to ensure specificity and robustness. For
instance, the development of a model for moisture should use spectral ranges at
1450 nm and or 1920 nm corresponding to the -OH absorption band. But the
overlap of chemical absorption bands in the NIR range usually means that large
variable ranges are included in the model.