This is a ready-to-use worksheet for setting and applying out-of-trend triggers on multivariate process analytical technology (PAT) data in continuous manufacturing (CM), where the unit of analysis is a moving window on a correlated data stream rather than a single batch time point. It pairs with the Worksheet: Stability OOT Limit Setting for the univariate case; use this one when the process is monitored by a multivariate model (PCA or PLS) on several correlated PAT or process variables. Replace every <<FILL: ...>> placeholder. A worked specimen follows. This worksheet describes standard multivariate SPC method in original wording; have a qualified statistician or chemometrician confirm the model for your data, and verify each cited regulation before you rely on it.
Worksheet header
| Field | Entry |
|---|---|
| Worksheet number | <<FILL: WS-ID>> |
| Process / unit operation | <<FILL: e.g. continuous direct-compression line, Train 2>> |
| Model type | PCA / PLS / Other: <<FILL>> |
| Variables included in the model | <<FILL: list, e.g. feed rate, blend NIR scores, compression force, tablet weight>> |
| Reference (training) data set | <<FILL: batches/campaigns, date range, n observations>> |
| Model software and version | <<FILL>> |
| Prepared by / date | <<FILL>> |
| Reviewed by (statistician/chemometrician, QA) / date | <<FILL>> |
1. Confirm the process is a genuine multivariate candidate
Do not build a multivariate model out of habit. Confirm first:
- At least
<<FILL: minimum, e.g. 3>>of the monitored variables are known to correlate under normal operation (check with existing process knowledge, DoE data, or a correlation matrix from historical data). - The process runs long enough, or continuously enough, that a moving-window view is more meaningful than a per-lot view.
<<FILL: describe the run pattern>> - A univariate control chart on each variable individually could plausibly miss a real excursion because no single variable crosses its own limit while the combination is abnormal.
<<FILL: brief rationale or reference to a known near-miss>>
If none of these hold, use univariate trending per the stability/trend OOT limit worksheet instead; a multivariate model adds complexity that only pays for itself when the variables genuinely move together.
2. Build and document the reference model
| Field | Entry |
|---|---|
| Reference data window and rationale for representativeness | <<FILL>> |
| Number of latent variables / components retained | <<FILL>> |
| Percent variance explained (cumulative) | <<FILL>> |
| Variables excluded from the model, and why | <<FILL: or "none">> |
| Known process states covered by the reference set (startup, steady state, planned rate changes, shutdown) | <<FILL>> |
| Model validation approach (cross-validation, hold-out set) | <<FILL>> |
State explicitly whether one model covers all process states or whether separate models are used per state; applying a steady-state model to a startup window is a common source of nuisance flags, not a real excursion.
3. Define the trigger statistics
| Statistic | What it detects | Limit basis | Value |
|---|---|---|---|
| Hotelling T-squared | An abnormal combination of otherwise-in-range variable values (within the model’s known correlation structure) | <<FILL: e.g. 95%/99% confidence limit from the F-distribution given N and the number of components>> | Alert <<FILL>>, Action <<FILL>> |
| Residual distance (Q-residual / DModX) | A mode of variation the model has never seen (an event outside the model’s known structure entirely) | <<FILL: e.g. chi-square approximation from the reference set residuals>> | Alert <<FILL>>, Action <<FILL>> |
| Contribution plot review threshold | Which variable(s) drove a flagged T-squared or residual event | <<FILL: e.g. review any variable contributing more than 20% of the statistic>> | <<FILL>> |
Set both an alert and an action level for each statistic if your program supports a graded response, the same logic as the univariate alert/action distinction. Record the limits and their basis before the model goes live in routine monitoring; a limit chosen after seeing an excursion is not defensible, the same principle as any other OOT limit.
4. Confirm the signal before chasing a process cause
A multivariate flag has a different confirmation sequence than a single-analyte OOT result. Work through this before treating the flag as a real process event.
| Check | Outcome | Notes |
|---|---|---|
| Sensor / probe health (fouling, calibration, optical alignment) | OK / Issue found | <<FILL>> |
| Known state transition in progress (startup, rate change, planned hold) at the flag time | Yes / No | <<FILL>> |
| Raw-material lot change since the model was last qualified | Yes / No | <<FILL>> |
| Model still within its qualified applicability domain (no material scale, formulation, or equipment change since last requalification) | Yes / No | <<FILL>> |
| Contribution plot reviewed: which variable(s) drove the statistic | <<FILL>> | |
| Reproducible on a repeat window, or a single isolated event | <<FILL>> |
If a sensor, model-applicability, or known-state cause explains the flag, document it, correct or exclude it under change control (original data preserved), and reassess. Do not silently drop flagged windows from the record; a pattern of “explained away” flags is itself a signal about model or sensor health that belongs in the model’s periodic requalification review.
5. Root cause and disposition (if confirmed real)
| Field | Entry |
|---|---|
| Root cause (process, material, equipment) | <<FILL>> |
| Material affected: in-process hold, downstream disposition | <<FILL>> |
| Was any material released before the flag was resolved; if so, retrospective assessment | <<FILL: or N/A>> |
| CAPA reference | <<FILL: number or N/A>> |
| Model update needed (retrain, add a state, exclude a variable) | <<FILL: or "no change">> |
Acceptance criteria
- The candidate variables genuinely correlate and a univariate approach was considered and ruled insufficient before building a multivariate model.
- T-squared and residual limits are defined and documented before the model runs in routine monitoring, with a stated basis, not reverse-engineered after a flag.
- Every flag is confirmed against sensor health, known state transitions, raw-material changes, and model applicability before a process root cause is pursued.
- A contribution-plot review identifies which variable(s) drove any flagged statistic; the investigation does not stop at “the model flagged something.”
- The model itself is on a defined periodic requalification cycle, with a trigger for an out-of-cycle requalification when raw materials, scale, or equipment change.
References
FDA Guidance for Industry, PAT, A Framework for Innovative Pharmaceutical Development, Manufacturing, and Quality Assurance (2004). ICH Q13, Continuous Manufacturing of Drug Substances and Drug Products (2022; adopted as FDA guidance 2023). ICH Q8(R2), Pharmaceutical Development, and ICH Q10, Pharmaceutical Quality System, for the process-understanding and monitoring expectations that PAT and CM build on. 21 CFR 211.110 (sampling and testing of in-process materials) and 211.180(e) (trend evaluation).
Confirm the current version and clause numbers of each reference before issue.
Revision history
| Version | Date | Author | Summary of change |
|---|---|---|---|
<<FILL: 1.0>> | <<FILL: date>> | <<FILL: author>> | Initial worksheet. |
Approvals
| Role | Name | Signature | Date |
|---|---|---|---|
| Prepared by | <<FILL>> | ||
| Statistician / chemometrician | <<FILL>> | ||
| Quality Assurance | <<FILL>> |
Filled specimen
Illustrative continuous direct-compression line, PLS model on six variables (feed rate, two blend NIR scores, compression force, tablet weight, ejection force), reference set built from three qualification campaigns.
| Field | Entry |
|---|---|
| Reference set | 3 qualification campaigns, steady-state windows only, n = 4,100 observations |
| Components retained | 3, cumulative variance explained 91% |
| T-squared limits | Alert at 95% confidence limit, Action at 99% confidence limit (F-distribution basis, 3 components, N = 4,100) |
| Residual (Q) limits | Alert and Action set from chi-square approximation on reference-set residuals |
| Flag event | Day 14, 03:40 to 03:55, T-squared exceeded the action limit for 12 consecutive minutes |
| Confirmation | Sensor health OK; no scheduled state transition; raw-material lot for the primary excipient changed 2 hours before the flag; model last qualified against the prior excipient lot |
| Contribution plot | One blend NIR score drove 68% of the T-squared statistic |
| Conclusion | Real signal, linked to the new excipient lot’s particle-size distribution shifting the blend NIR signature outside the model’s known range; not a sensor fault |
| Material disposition | In-process hold on affected output; retrospective testing confirmed product quality within specification; released with enhanced sampling for the remainder of the campaign |
| Model action | Excipient lot range added to model qualification scope; supplier notified; CAPA opened to tighten incoming particle-size specification or expand the model’s qualified range, decision pending |
The point of the specimen: the flag was not treated as automatic proof of a process problem, and it was not waved away either. The contribution plot pointed at one variable, the raw-material change explained it, in-process material was held pending confirmation rather than released on faith, and the model’s own qualified range, not just the process, was identified as something that needed to change.
Common inspection findings this worksheet prevents
- A multivariate model running with no documented basis for its T-squared and residual limits, so “why is the action limit set there” has no answer.
- A flagged event dismissed as “just the model” with no contribution-plot review and no check of sensor health, state, or raw-material changes.
- A model applied outside its qualified range (a new raw-material lot, a scale change) with no requalification trigger defined.
- Univariate control charts on each PAT variable used as if they were sufficient, when the variables are known to correlate and a real excursion could hide inside individually-normal-looking values.
- Material released before a flagged event was resolved, with no retrospective assessment documented.
How to adapt this worksheet
- Set your worksheet numbering and confirm with your process and statistics team that a multivariate approach is warranted for this process before building the model.
- Replace the variable list, model type, and limit basis with your own validated tool’s output; this worksheet records the inputs and results, it does not replace a validated chemometric platform.
- Align the confirmation checks in section 4 with your actual PAT sensor maintenance and model lifecycle management procedures.
- Tie the requalification trigger to your change-control process so a raw-material, scale, or equipment change automatically prompts a model review.
- Confirm every reference before issue, and describe any compendial or ICH guidance in your own words rather than reproducing its text.