How to use this document
Copy the template below into your quality system. Replace every <<FILL: ...>> placeholder with your company, system, and process specifics, then route it through your own document control, review, and approval. The filled specimen shows one worked example; adapt the numbers and scope to your operation and verify each cited regulation or standard against the current source before you rely on it.
This is a generic, originally worded starting point. It contains no confidential, proprietary, or employer material, it is drafted with AI assistance and reviewed, it is not legal or compliance advice, and adopting it does not by itself create or prove compliance. See the Disclaimer and Terms.
This is a ready-to-use plan for building and operating a multivariate Continued Process Verification (CPV) monitoring model, for a process with enough correlated critical process parameters and quality attributes that univariate charts alone leave a real gap. It is a companion to the univariate CPV plan, not a replacement for it; univariate charts stay in place, and this plan covers what the parameter traceability matrix’s multivariate section only flags as in use or not. Replace every <<FILL: ...>> placeholder with your own specifics and route it through document control. A filled specimen follows. Verify each cited reference against the current source before you rely on it.
1. Purpose
This plan defines how a multivariate monitoring model for <<FILL: parameter group>> is built, qualified, operated, and maintained, so a correlated shift across several parameters that stays inside every individual control limit is still detectable, and so the model itself is never a less governed part of the CPV program than the univariate charts sitting alongside it.
2. Scope
3. Risk basis for building a multivariate model
State why a univariate program alone leaves a real gap here, not as a generic statement. This section should identify the specific correlated-parameter group, the shared upstream driver it reflects, and the batch history available to build from.
4. Approach
5. Training set qualification
This is the most consequential step in the plan. A model trained on a population that quietly includes an unresolved deviation, or batches spanning an unflagged raw material change, will learn that irregularity as normal and fail to flag a real recurrence of it later.
6. Model build and validation
7. Ongoing operation
8. Retraining triggers
A retraining never follows directly from an inconvenient flag. State the predefined triggers below; a retraining with no trigger checked is not eligible for approval.
9. Model change control
Any retraining, variable-set change, or threshold change runs through change control with the same discipline a univariate control limit recalculation requires. A retraining that quietly absorbs the batches behind a recent flag into the next model version, with no independent trigger from section 8, is not approved under this plan.
10. Roles and responsibilities
11. Schedule for initial build
12. Acceptance criteria for go-live
The model is eligible to go live once the training set is qualified per section 5, the model’s thresholds are documented with their statistical basis per section 4, at least one validation check per section 6 has been completed, and QA has approved this plan including the retraining triggers in section 8. A model running against live batches with no approved plan covering these points is not in a qualified state regardless of how long it has been in informal use.
13. References
PDA Technical Report No. 60 (Revised 2026), Process Validation: A Lifecycle Approach.
FDA guidance, “Process Validation: General Principles and Practices” (January 2011), Stage 3.
See the paired article, continued process verification: building a Stage 3 program that actually monitors, for the reasoning behind operationalizing multivariate monitoring as a written plan.
Confirm the current version of each reference before issue.
14. Revision history
15. Approvals
Filled specimen
Extract for an example allogeneic cell therapy drug substance process. Company and numbers are illustrative.
At go-live, the model correctly flagged, in retrospective scoring, both of the historically low-viability, high-yield batches with a T-squared value above the 95% threshold, the same pattern that passed univariate review at the time; this retrospective check was documented as the section 6 validation basis before the model went live on new batches.
Common inspection findings this plan prevents
- A multivariate model running against live batches with no written plan, discoverable only by asking who built it and finding the answer is a slide deck, not a governed document.
- A training set that was never checked against open deviations or unconfirmed changes, so the model’s definition of “normal” quietly includes an unresolved irregularity.
- A retraining that followed immediately after a flag, with no independent trigger from a predefined schedule or batch count, effectively erasing the flag from the model’s memory.
- No distinction documented between a T-squared flag and an SPE flag, so a reviewer cannot tell whether a batch fits a known pattern at an unusual intensity or represents a pattern the model has never seen.
How to adapt this plan
- Build section 3’s risk basis from an actual correlation analysis on your own historical data, not an assumed relationship between parameters that happen to share a process step.
- Set the training set exclusion criteria in section 5 before looking at which batches would pass or fail them; deciding exclusions after seeing the model’s output defeats the qualification step’s purpose.
- Set retraining triggers in section 8 that match your own CPV plan’s philosophy on univariate limit recalculation; the two should not use a looser standard for the harder-to-audit multivariate side.
- Confirm every reference in section 13 against the current published version before issue.
