12 Method Development
285
As a result, some risk will be passed onto the method users because model robustness needs to be monitored throughout lifecycle management. They need to accept
risk factors that may arise from previously unseen sources of variability as well as
factors that were simply not identified a priori. These unforeseen sources of variability and their potential impacts on the method will be handled through the method
lifecycle management process and a plan should be in place to maintain robustness
of the model throughout its lifecycle.
12.3.5 Method Development
When reaching this stage of development, a significant amount of information is
known about the desired performance criteria (analyte target profile), the suitability of
the sampling approach and analytical technique (feasibility study), and the factors that
may affect the model (risk assessment). These learnings are now taken into account to
optimize the sampling methodology, the sample presentation to the analytical instrument, the data collection (to ensure measurement representativity), etc. It is essential
that these factors are set prior to spectral data collection for the calibration, test and
if applicable validation sets. Any change to instrument configuration and collection
parameters should be assessed for impact on the method. This is particularly important in cases where the calibration set is collected with a different experimental setup,
e.g., off-line calibration for in-line use. Such an example may occur when in-line
calibration samples cannot be prepared and presented in a representative manner.
12.3.5.1 Sample Set Membership Considerations
The sample membership of the calibration set, test set(s), and potential validation set
should be carefully designed. Specifically, the validation set should be completely
independent from the calibration and test sets, should represent expected variability
from the process, and should challenge the model across an appropriate range that
the model is intended to cover. The test set(s) should be representative of variability
expected during normal operating conditions. The calibration set should be built
from the experimental variables identified in the risk assessment. If sample particle
size, density, batch-to-batch or seasonal variability, etc., is expected to change, that
information should be included in the model. If too much variability is included,
the model may suffer from a lack of accuracy at the expense of robustness. If the
process is expected to be highly variable, local chemometric methods segmenting
large ranges of variability into smaller segments for model development may be
better suited [12].
The model should be qualified for its use irrespective of the application, and the
requirements for qualification (or validation) will vary across industries. In the pharmaceutical industry, it is necessary to satisfy ICH Q2(R1) requirements for method
validation in addition to the relevant guidelines (e.g., EMA and FDA guidelines on
285
As a result, some risk will be passed onto the method users because model robustness needs to be monitored throughout lifecycle management. They need to accept
risk factors that may arise from previously unseen sources of variability as well as
factors that were simply not identified a priori. These unforeseen sources of variability and their potential impacts on the method will be handled through the method
lifecycle management process and a plan should be in place to maintain robustness
of the model throughout its lifecycle.
12.3.5 Method Development
When reaching this stage of development, a significant amount of information is
known about the desired performance criteria (analyte target profile), the suitability of
the sampling approach and analytical technique (feasibility study), and the factors that
may affect the model (risk assessment). These learnings are now taken into account to
optimize the sampling methodology, the sample presentation to the analytical instrument, the data collection (to ensure measurement representativity), etc. It is essential
that these factors are set prior to spectral data collection for the calibration, test and
if applicable validation sets. Any change to instrument configuration and collection
parameters should be assessed for impact on the method. This is particularly important in cases where the calibration set is collected with a different experimental setup,
e.g., off-line calibration for in-line use. Such an example may occur when in-line
calibration samples cannot be prepared and presented in a representative manner.
12.3.5.1 Sample Set Membership Considerations
The sample membership of the calibration set, test set(s), and potential validation set
should be carefully designed. Specifically, the validation set should be completely
independent from the calibration and test sets, should represent expected variability
from the process, and should challenge the model across an appropriate range that
the model is intended to cover. The test set(s) should be representative of variability
expected during normal operating conditions. The calibration set should be built
from the experimental variables identified in the risk assessment. If sample particle
size, density, batch-to-batch or seasonal variability, etc., is expected to change, that
information should be included in the model. If too much variability is included,
the model may suffer from a lack of accuracy at the expense of robustness. If the
process is expected to be highly variable, local chemometric methods segmenting
large ranges of variability into smaller segments for model development may be
better suited [12].
The model should be qualified for its use irrespective of the application, and the
requirements for qualification (or validation) will vary across industries. In the pharmaceutical industry, it is necessary to satisfy ICH Q2(R1) requirements for method
validation in addition to the relevant guidelines (e.g., EMA and FDA guidelines on