The second revision of the International Council for Harmonisation (ICH) guideline on Analytical Procedure Validation, Q2(R2), was adopted by the Committee for Medicinal Products for Human Use (CHMP) on 14 December 2023 and became legally effective in the European Union on 14 June 2024. For anyone working with near-infrared (NIR) or Raman measurement in a regulated process, it matters, because it is the first version of Q2 to talk about multivariate calibration. Q2(R1) simply did not contemplate a spectrum with several hundred wavelength variables feeding a partial least squares (PLS) model.
What is striking, once you read it, is how short that treatment is. It is comprised of Section 2.5 (Considerations for multivariate analytical procedures) that runs to roughly a page, and includes some paragraphs scattered through the range (sec. 3.2) and accuracy and precision (sec. 3.3) sections, plus one worked example for an NIR core tablet assay in Annex 2. That is a deliberate scope, not an omission, as Q2(R2) sets out the structure of the validation argument, and leaves the specifics of performance criteria to the applicant. This is reasonable since the specifics vary wildly with the intended purpose and the technology used. However, this places real responsibility on the team building the case, and it's worthwhile discussing the points where judgment is most needed.
What the guideline specifies
Q2(R2) describes multivariate development and validation as two phases. In the first, calibration data build the model and test data are used for internal testing and optimisation, including selection of the number of latent variables. That test data may be a separate set or part of the calibration set used rotationally. In the second phase, an independent validation set of samples challenges the finished model. Note that the independent set of samples is defined as samples that are not included in the calibration set but may originate from the same batch. Finally, for identification libraries, validation includes challenging the library with samples it does not contain to assess their discriminative ability.
Calibration and validation samples need reference values, normally from a validated or pharmacopoeial procedure, and the reference procedure's performance should equal or exceed what you expect from the multivariate procedure. Reference analysis and spectral collection should happen on the same samples within a sensible time window. To demonstrate accuracy (section 3.3.1.4), the guideline recommends using the root mean-squared error of prediction (RMSEP), and later (sec. 3.3.2.4) notes that RMSEP encompasses both accuracy and precision. For response across the range (section 3.2.2.3), it calls for the relationship between predicted and reference values, plus information on how model error behaves across the calibration range, typically through residual plots.
While this is a coherent framework, it is also a framework that leaves the setting of thresholds to the applicant.
The one benchmark Q2(R2) offers needs careful consideration
Q2(R2) gives univariate analysts several concrete anchors. A signal-to-noise ratio of 3:1 for detection limit and at least 10:1 for quantitation limit. A minimum of five concentrations across the range for linearity. Nine determinations, or six at the 100 percent level, for repeatability. Relative response factors outside 0.8 to 1.2 needing a correction factor. For multivariate procedures there is only one benchmark, and it is soft: If RMSEP is comparable to an acceptable root mean-squared error of calibration (RMSEC), the model is indicated to be sufficiently accurate against an independent test set.
The RMSEP-RMSEC comparison criterion is genuinely useful as it tests whether the model generalises from calibration to independent samples. However, it does not test whether either figure is good enough for the decision the measurement supports. An overfitted model with an unimpressive RMSEC can produce a perfectly comparable RMSEP and still be unfit. The threshold must come from the Analytical Target Profile (ATP) described in ICH Q14 (Analytical Procedure Development), calculated from the specification and the role the measurement plays in the control strategy. Importantly, this threshold needs to be defined before the model is built, i.e. during method development, not after you see what the model achieves.
Reference error deserves its own characterisation before calibration begins
Q2(R2) notes that multivariate accuracy depends on the distribution of calibration samples and on the error of the reference procedure. It also requires reference performance to meet or exceed the expected performance of the multivariate procedure. These are reasonable requirements, but the guideline doesn't provide a method for separating the two contributions once they're combined in a single RMSEP figure.
This is a recognised limitation in chemometric method development: model accuracy cannot exceed the precision of the reference method it was calibrated against, since reference error becomes part of the model's own error. When a model's performance plateaus despite further calibration work, it's worth checking whether the constraint is the reference method rather than the model itself, since no amount of additional modelling will resolve an error that originates in the reference data.
The practical step is to characterise reference precision before calibration and state the resulting floor on RMSEP in the protocol. If the floor sits above your ATP, the answer is not a better model. It is a better reference procedure, or a different measurement strategy, and it is much cheaper to learn that during the feasibility/development phase.
Where Q14 supplies what Q2(R2) leaves open
Both points discussed above resolve the same way: the number that would settle them, an acceptance threshold for RMSEP, and a ceiling on acceptable reference error, isn't something Q2(R2) supplies. Rather, it comes from the ATP that is set in the process development phase (which is pure ICH Q14 territory), long before the validation protocol is written.
Done well, this means the validation study becomes a confirmation of decisions already made and justified, rather than the place where they're worked out for the first time. Q2(R2) tells you what argument the validation study needs to make. Q14 is where the terms of that argument, i.e., the thresholds and the error budget, are meant to be set. Building that case and being able to defend it in front of an assessor remains the applicant's work either way.
Amalia Technologies supports life sciences organisations across the PAT lifecycle, from feasibility analysis and design of experiments through chemometrics and model validation, alongside the software validation and data integrity work that PAT deployments depend on. If you are preparing a multivariate procedure for validation, or reassessing one already in routine use, our PAT team can help you work out where the evidence is thin before an assessor does.
