12 Method Development
291
multivariate regression models are still valid, diagnostic metrics need to be routinely
collected and trended. The calculation of diagnostic values such as the Hotelling’s
T
2 and the Q-residuals will confirm for every spectrum that it appropriately belongs
to the model space. This will provide confidence that the spectrometer, sample, and
environment are all producing a spectrum that is as required for the model to perform
as developed, tested, and potentially validated.
Tracking and trending of the diagnostic values are essential to the successful
deployment of a NIRS-based method. Statistical limits can then be set to raise
alarms when a measurement appears to be different or tending toward process limits.
The limits for the diagnostics can be set by incorporating the assessment into the
method qualification or in parallel with initial batch experience during the first several
production runs. These diagnostics will determine whether the sample belongs to the
model space (Hotelling’s T
2 ) and whether there is excessive unmodelled variance
(Q-residual). Note that it is possible for a prediction to be within process limits but
for the diagnostics to flag the sample as being “different” or an outlier. If a sample
is determined as exceeding an outlier limit, it should not be assumed however that
the sample is different or that the instrument is not functioning properly. If justified
by consistent out of specification results beyond the expected statistical limits, an
investigation of the system, the process, and the materials used should be undergone
to determine the root cause. If it is determined that the sample was appropriate and
that no instrument issue occurred, an update to the method may be considered.
12.4 Method Lifecycle
Maintaining method performance years after method development is the current
challenge that NIRS practitioners are facing to ensure the successful use of NIRS,
regardless of the field of application. A number of scenarios can be envisioned that
may affect a method such as instrument changes and replacement (either caused by
optical part replacement, instrument failure or transfer of the process to a new manufacturing site), process changes (process wear and tear and process improvements),
and raw materials changes (long-term lot-to-lot variability and sourcing changes).
It is also important to systematically monitor risk factors identified during the risk
assessment until they are deprioritized or eliminated. Alternatively, when new risks
arise, it may be necessary to either control the risks or update the model to handle that
new variability. Finally, periodic assessment against the reference method should be
performed to demonstrate the method is still performing adequately. This may be
done even if predictions and diagnostics do not show any new concerns. In a regulated environment such as the pharmaceutical industry, it shall be noted that periodic
assessment does not mean periodically performing a verification exercise against the
original validation dataset, i.e., re-validation. Re-validation is only needed when a
model update takes place.
If it is determined that an update to the model is required, the following approaches
could be used:
291
multivariate regression models are still valid, diagnostic metrics need to be routinely
collected and trended. The calculation of diagnostic values such as the Hotelling’s
T
2 and the Q-residuals will confirm for every spectrum that it appropriately belongs
to the model space. This will provide confidence that the spectrometer, sample, and
environment are all producing a spectrum that is as required for the model to perform
as developed, tested, and potentially validated.
Tracking and trending of the diagnostic values are essential to the successful
deployment of a NIRS-based method. Statistical limits can then be set to raise
alarms when a measurement appears to be different or tending toward process limits.
The limits for the diagnostics can be set by incorporating the assessment into the
method qualification or in parallel with initial batch experience during the first several
production runs. These diagnostics will determine whether the sample belongs to the
model space (Hotelling’s T
2 ) and whether there is excessive unmodelled variance
(Q-residual). Note that it is possible for a prediction to be within process limits but
for the diagnostics to flag the sample as being “different” or an outlier. If a sample
is determined as exceeding an outlier limit, it should not be assumed however that
the sample is different or that the instrument is not functioning properly. If justified
by consistent out of specification results beyond the expected statistical limits, an
investigation of the system, the process, and the materials used should be undergone
to determine the root cause. If it is determined that the sample was appropriate and
that no instrument issue occurred, an update to the method may be considered.
12.4 Method Lifecycle
Maintaining method performance years after method development is the current
challenge that NIRS practitioners are facing to ensure the successful use of NIRS,
regardless of the field of application. A number of scenarios can be envisioned that
may affect a method such as instrument changes and replacement (either caused by
optical part replacement, instrument failure or transfer of the process to a new manufacturing site), process changes (process wear and tear and process improvements),
and raw materials changes (long-term lot-to-lot variability and sourcing changes).
It is also important to systematically monitor risk factors identified during the risk
assessment until they are deprioritized or eliminated. Alternatively, when new risks
arise, it may be necessary to either control the risks or update the model to handle that
new variability. Finally, periodic assessment against the reference method should be
performed to demonstrate the method is still performing adequately. This may be
done even if predictions and diagnostics do not show any new concerns. In a regulated environment such as the pharmaceutical industry, it shall be noted that periodic
assessment does not mean periodically performing a verification exercise against the
original validation dataset, i.e., re-validation. Re-validation is only needed when a
model update takes place.
If it is determined that an update to the model is required, the following approaches
could be used: