A process always varies. The useful question is not whether it varies, but which of its variations reach the product — and by what route.
PAT-INDUSTRY establishes those links with process and development teams, in pharma, biopharma, chemicals, cosmetics and food. This is the step that decides whether a measurement will control or only observe.
Where the variation comes from
What comes in
- Raw materials: batch, supplier, water content on receipt, particle size
- Excipients and processing aids, whose variability is rarely tracked
- Ambient conditions: season, humidity in the workshop
These sources are endured. A measurement at the inlet makes them visible before they propagate.
What gets set
- Operating parameters: temperature, flow, speed, duration, pressure
- Equipment: wear, fouling, format change
- Practices: shift, station, procedure as interpreted
These sources can be controlled. Provided you know which ones matter.
The vocabulary of the dossier, and why to use it
Four acronyms run through every pharmaceutical dossier, and they do not mean the same thing. Confusing them costs time in the room.
| Acronym | What it is | Example |
|---|---|---|
| QTPP | The target profile of the product: what it has to do, before any consideration of process | Tablet, immediate release, target dose and stability |
| CQA | An attribute of the product whose value must stay within a range | Content uniformity, dissolution, water content |
| CPP | A process setting whose variability reaches a CQA | Impeller speed, inlet air temperature, main compression force |
| CMA | A property of an incoming material that reaches a CQA | Particle size of an excipient, water content of a lot |
The chain reads in one direction, and that is what makes it useful: the CMAs and CPPs of one operation produce the attributes of the intermediate, which become the CMAs of the next. That is why an inlet variation turns up three operations downstream, and why it is rarely looked for where it started.
Criticality is not a property of the parameter
It is a property of the relationship between that parameter and an attribute of the product. The same temperature setting is critical on one formulation and has no effect on the next. A table of critical parameters valid for a plant does not exist: it is valid for one product on one line, and it reopens when either one changes.
ICH Q8(R2) defines a critical process parameter as one whose variability has an impact on a critical quality attribute, and which should therefore be monitored or controlled. The definition sets no threshold: an impact is enough.
So it says what is critical, not where to place a measurement. What decides the measurement is impact crossed with real variability: two highly influential parameters that never move in your plant call for no measurement; a moderately influential but very unstable one does.
Four levels rather than two
“An impact is enough” is a definition that sorts everything into two boxes. In practice the “critical” box becomes the monitoring plan, and it fills up with parameters whose real effects bear no comparison to one another.
A technical guide published by USP in 2023 proposes a graduation, and states along the way the criticism few texts write down: a binary approach is limiting, because the wording of ICH Q8 — “whose variability has an impact” — is imprecise.
| Level | What happens when the parameter drifts within its own limits |
|---|---|
| High impact | The CQA goes outside its limits |
| Medium impact | The CQA shifts measurably, but stays within limits |
| Low impact | Minor shift, but measurable |
| No impact | No measurable shift, or no mechanistic relationship |
The graduation changes what you do, not just what you write. A high-impact parameter calls for control: a loop, a specification, an endpoint criterion. A medium-impact parameter calls for surveillance — you watch it, you document it, you do not necessarily commit an in-line measurement to it. That is how a monitoring plan stays proportionate.
And the last row deserves its own comment. “No mechanistic relationship” is the only one of the four that is established by reasoning rather than by trial: if nothing in the physics or chemistry of the operation links the parameter to the attribute, an observed absence of correlation is no longer a coincidence to be checked. It is the one case that can be closed without an experiment.
The IPO matrix, and what it brings to the surface
The most useful tool of this step is also the simplest. For each unit operation you write three columns: what comes in, what gets set, what comes out.
Filled in honestly, the matrix surfaces three things a discussion does not. Untracked inputs — a raw material whose variability nobody measures although it runs through the whole process. Unmeasured outputs — an attribute assumed under control because it has never raised a question. Assumed links — “everyone knows that one matters”, with no data behind it.
It is built with your experts, in session. We bring the frame and the questions; the content comes from the people who run the process, operators included, who often know before anyone else what is drifting.
Correlation, causation, and the trap of the model that works
This is the most expensive watch-point, and it does not show in the numbers.
A statistical model can link two quantities without either acting on the other: both depend on a third one, absent from the model. The model is excellent on historical data, and it becomes wrong the day that third quantity changes behaviour — new supplier, new season, new equipment.
Controlling on a correlative link produces a loop that moves without improving the product, and that fails without warning. The remedy is not statistical, it is experimental: vary the suspect parameter deliberately, within an acceptable range, and see whether the attribute follows. A design of experiments, even a reduced one, settles what no amount of historical data will settle.
It is also what the FDA PAT guidance asks for when it defines a well-understood process: critical sources of variability identified and explained. The second word is not decoration.
Frequently asked
Can production history replace trials?
Partly. History shows what varied and what followed; it does not show what would have followed had you moved a parameter that stayed fixed. It is excellent at naming suspects, insufficient to convict them. The usual combination is history that points, then a few targeted trials that conclude.
How many critical parameters are usually found?
Fewer than feared. On a unit operation the initial list often holds twenty to thirty candidates; rarely more than three or four survive the analysis. That is good news for the project: three parameters to follow is a realistic control strategy.
Does this work have to be redone when the process changes?
Not entirely, but it has to be reopened. A change of supplier, equipment or scale shifts the ranges of variation, and therefore criticality. A parameter that is not critical at small scale can become critical in production. It is one of the reasons process knowledge is maintained rather than archived.
Which drift keeps coming back without an explanation?
Describe it, with the operation concerned. Forty-five minutes are enough to draw a first matrix and see which leads deserve a measurement.