This is a field-notes piece that synthesizes patterns from multiple practitioners who have hosted FDA or EU competent-authority inspections in which process analytical technology and chemometric models were in scope. No quote is attributed to a single named source; every observation is a composite of multiple interlocutors and is presented as such. Where a pattern aligns with a published guidance, we cite the guidance.
We use this format because pharmaceutical inspection commentary is unusually politically charged. People who have lived through an inspection rarely want to be quoted on what an inspector said or did, and inspectors themselves rarely speak on the record outside agency-sanctioned channels. The composite removes the personal attribution and leaves the recurring patterns visible.
Pattern 1: inspectors do not arrive looking for the chemometric model
The most consistent observation. PAT systems are not the inspection’s starting point, even on inspections where PAT is in scope. The inspector arrives with a process - a product, a manufacturing line, a release decision - and follows the audit trail of that process to whatever measurement supports it. If a chemometric model produces a release-relevant number, the model becomes part of the inspection by inheritance, not because the inspector picked it out of the QMS.
The practical implication: a site that has prepared a thick binder titled “PAT validation” and a thin set of procedures around how the PAT system is operated on the floor has prepared for the wrong inspection. The inspector will skim the binder and spend the next two days on the operations side. Sites that have inverted this - lean qualification record, detailed operating and lifecycle procedures - report easier inspections.
Pattern 2: the question is “what does this number mean, and how do you know”
Practitioners describe a recurring sequence. The inspector picks a batch record. The batch record references a release-relevant number. The number came from a chemometric model. The inspector asks, in some form: what does this number mean, how was the model that produced it built, how do you know it is still working, and what would you do if it stopped working.
The four questions are independent, and each has a defensible answer in the published guidance. What it means lives in the analytical procedure description per ICH Q2(R2) and Q14. How it was built lives in the model development record per Q14 and, indirectly, in the PAT framework. How you know it is still working lives in the ongoing performance verification programme. What you would do if it stopped working lives in the change-control and deviation procedures.
Sites that can walk the inspector through these four documents in fifteen minutes, without flipping between binders, finish the conversation. Sites that cannot - because the four answers live in four different parts of the QMS, owned by four different functions, and nobody has rehearsed the sequence - spend a day on what should have been a fifteen-minute exchange.
Pattern 3: the deviation log is where the inspection is won or lost
Echoing what validation engineers describe, inspectors spend disproportionate time on the deviation log. For a PAT system, the deviation log is where the practical reality of the model lives: drift events, recalibration events, failed performance-verification samples, instrument repairs that required model re-verification, transfers between probes.
The patterns practitioners describe:
- A deviation log with no PAT-related entries over a year of operation is a red flag, not a green one. Either the system is not being operated as documented, or the thresholds are set so loosely that real drift is being ignored.
- A log full of small, self-reported, closed events with documented root causes reads as a site that is paying attention.
- A deviation closed with “operator retrained” and no system change is read sceptically. Human-error closures are often a euphemism for an unresolved systemic issue.
The implication is that the deviation procedure for the analyzer should capture small events, not just batch-impacting events. A periodic-performance-verification sample at the edge of the acceptance window should generate a logged event with a documented assessment, even if no batch is affected.
Pattern 4: data integrity is read through the audit trail
Data integrity is the area where inspector expectations have hardened the most over the last decade. The FDA guidance on data integrity and EU Annex 11 set the framework. For a chemometric model running in a production environment, the practical reading inspectors apply is:
- The raw spectrum is the original record. It must be retained, time-stamped, and attributable to a batch and a measurement event.
- The model prediction is a derived record. The model version that produced the prediction must be traceable from the prediction.
- The audit trail for the model itself - which version was deployed when, who approved the deployment, what changed between versions - must be reviewable independently of the IT system that runs the model.
- 21 CFR Part 11 applies to the prediction record the same way it applies to any other GMP-relevant electronic record. Electronic signatures on model approvals are inspected the same way as electronic signatures on batch records.
Sites that store the raw spectrum, the model version, and the prediction in a single linked record per measurement do well. Sites that store the prediction in the historian and the raw spectrum in a separate vendor system without a foreign key between them spend the inspection reconstructing the link.
Pattern 5: the inspector who knows chemometrics is rare, and that is not a defence
A myth that circulates in the PAT community is that inspectors do not know chemometrics, so the model details can be hand-waved. Practitioners who have been inspected describe the opposite pattern: the inspector does not need to know chemometrics in detail because the inspection is read against the company’s own procedures and against the regulatory framework, neither of which require the inspector to recalculate a PLS regression.
What the inspector does need to do is read the company’s analytical procedure document and confirm that the model is being operated as that document describes. That confirmation does not require chemometric expertise; it requires the company’s document to be specific enough to be checked against the operating record. Vague analytical-procedure documents that describe the model as “a multivariate calibration that predicts assay” without specifying the algorithm, the wavelength range, the preprocessing, and the acceptance criteria are read as a documentation problem regardless of how good the underlying model is.
The defence against an inspector who does not know chemometrics is not vagueness. It is a procedure specific enough that any competent reader can check it against the operating record.
What practitioners do not describe as the failure mode
A few items that did not recur in the field notes as inspection failure modes:
- The regulatory framework itself. Between the FDA PAT guidance, ICH Q2(R2), ICH Q14, the process validation guidance, the data integrity guidance, Annex 11, and Annex 15, the framework for inspecting a PAT system is in place. Practitioners did not describe gaps in the framework as the source of inspection findings.
- Vendor documentation. Major analyzer vendors now ship FAT and DQ documentation that practitioners describe as adequate. Findings are rarely written against the vendor’s documents.
- Model performance. Models that are operating outside their acceptance window are caught by the company’s own monitoring before the inspector finds them, in nearly every case practitioners described. The inspection finding is not “the model is wrong” - it is “the company did not document what it did when the model drifted.”
The pattern that emerges across the conversations is consistent. Inspectors are not auditing the science of chemometrics. They are auditing whether the company is operating the system it said it would operate, whether it notices when the system deviates from its qualified envelope, and whether the documented response to deviation is the response that actually occurred. The companies that have built their PAT lifecycle around those three questions, rather than around producing a thick validation binder, are the ones who describe inspections as routine rather than as crises.