402
B. Igne and E. W. Ciurczak
associated with blend quality are homogeneity and stability. Blending is used in all
drug product manufacturing operations and has thus been the subject of extensive
research.
Batch blending
The traditional approach to batch blend quality evaluation relies on powder sampling
at various locations of the blender bin after a given time of mixing and analysis by
HPLC for the API content. It is a discrete operation that provides an evaluation of
an entire bin at a particular time point. But because it requires manual sampling, it
has been shown that it can induce errors [42]. Alternatively, a wireless NIR unit can
be mounted on a blender. While the amount of powder analyzed at each rotation is
usually lower than what would be considered by manual sampling (between 10 and
50 mg based on the collection optics and powder density), it allows the understanding
of blend kinetics for not only the active ingredient but also the major excipients, the
comparison of blend trends across batches without having to manually sample, and
the possibility to determine a blend end-point.
Blend analysis by NIRS has been performed off or at-line on sampled powders by
single point spectrometer [43] or imaging [44] but most of the work has been done
on-line with qualitative and quantitative methodologies. Hailey et al. and Sekulic
et al. first reported the use of on-line NIR with qualitative approaches to evaluate the end-point [45]. They used a moving block standard deviation to determine
when the blend was done evolving. In that approach, the pooled standard deviation across the variables is calculated for a set block of spectra and compared with
previous and subsequent blocks. Figure 18.5 shows the application of the moving
block standard deviation approach with a block size of 25 spectra onto NIR spectral
data collected on-line for blend comprised of acetaminophen (34.5%w/w), lactose
(34.7%/w), microcrystalline cellulose (24.8%/w), croscarmellose sodium (5.5%/w),
and magnesium stearate (0.5%/w) [46]. The spectra were preprocessed with Standard
Normal Variate to reduce the effects of physical properties. The approach determined
that the blend reaches a plateau of variability after about 40 rotations with no longterm trend. A threshold could be set using historical information to identify when
a blend has reached homogeneity. However, this approach requires historical information of what should be considered homogeneous. To address this limitation, an
F-Test-based method was proposed [47]. The spectral variance of two sequential
blocks is calculated and compared with an F-critical value calculated based on the
size of the blocks and the confidence limit. This approach avoids having to set a value
of standard deviation corresponding to blend stability but rather uses the block-toblock information to determine when the two populations have a similar variance,
indicating that the blend is no longer changing given the confidence limit considered.
These two methods use the spectral variance in the form of a spectrum standard
deviation. It could be envisioned that for a same resulting standard deviation, different
regions of the spectra could be changing; thus, indicating different phenomena taking
B. Igne and E. W. Ciurczak
associated with blend quality are homogeneity and stability. Blending is used in all
drug product manufacturing operations and has thus been the subject of extensive
research.
Batch blending
The traditional approach to batch blend quality evaluation relies on powder sampling
at various locations of the blender bin after a given time of mixing and analysis by
HPLC for the API content. It is a discrete operation that provides an evaluation of
an entire bin at a particular time point. But because it requires manual sampling, it
has been shown that it can induce errors [42]. Alternatively, a wireless NIR unit can
be mounted on a blender. While the amount of powder analyzed at each rotation is
usually lower than what would be considered by manual sampling (between 10 and
50 mg based on the collection optics and powder density), it allows the understanding
of blend kinetics for not only the active ingredient but also the major excipients, the
comparison of blend trends across batches without having to manually sample, and
the possibility to determine a blend end-point.
Blend analysis by NIRS has been performed off or at-line on sampled powders by
single point spectrometer [43] or imaging [44] but most of the work has been done
on-line with qualitative and quantitative methodologies. Hailey et al. and Sekulic
et al. first reported the use of on-line NIR with qualitative approaches to evaluate the end-point [45]. They used a moving block standard deviation to determine
when the blend was done evolving. In that approach, the pooled standard deviation across the variables is calculated for a set block of spectra and compared with
previous and subsequent blocks. Figure 18.5 shows the application of the moving
block standard deviation approach with a block size of 25 spectra onto NIR spectral
data collected on-line for blend comprised of acetaminophen (34.5%w/w), lactose
(34.7%/w), microcrystalline cellulose (24.8%/w), croscarmellose sodium (5.5%/w),
and magnesium stearate (0.5%/w) [46]. The spectra were preprocessed with Standard
Normal Variate to reduce the effects of physical properties. The approach determined
that the blend reaches a plateau of variability after about 40 rotations with no longterm trend. A threshold could be set using historical information to identify when
a blend has reached homogeneity. However, this approach requires historical information of what should be considered homogeneous. To address this limitation, an
F-Test-based method was proposed [47]. The spectral variance of two sequential
blocks is calculated and compared with an F-critical value calculated based on the
size of the blocks and the confidence limit. This approach avoids having to set a value
of standard deviation corresponding to blend stability but rather uses the block-toblock information to determine when the two populations have a similar variance,
indicating that the blend is no longer changing given the confidence limit considered.
These two methods use the spectral variance in the form of a spectrum standard
deviation. It could be envisioned that for a same resulting standard deviation, different
regions of the spectra could be changing; thus, indicating different phenomena taking