After performance qualification (PQ), Confiancapharmazon’s AVI guidance holds that any change in optics, software or handling systems requires requalification. Keeping a vision inspection system qualified means putting the model, the system and the process under change control together: any change is documented and evaluated to decide whether retesting is needed. EU GMP Annex 15 defines change control as the formal system in which qualified people from the relevant disciplines review proposed or actual changes that might affect validated status, so that the system stays in a validated state.

TechnoLynx defines and documents which parameters make up the qualified baseline (algorithm version, thresholds, optics and lighting settings). TechnoLynx also builds automated configuration logging and drift detection against the PQ baseline, and designs the requalification trigger logic with your quality and regulatory stakeholders. Each change still needs quality unit approval before implementation.

Why optics, lighting and the algorithm cannot be qualified separately

In our engineering view, a vision algorithm does not judge a vial directly. It judges the image that the camera, lens and lighting produced, so every threshold set at PQ is tuned to that particular image formation. Swap the lens or let the lamp dim, and the same threshold is now applied to a different distribution of pixel values. The algorithm’s code hasn’t changed, but its behaviour has.

The EU’s draft Annex 22 on artificial intelligence extends change control to the model’s physical inputs. It says the tested model, the system it runs in and the whole process it automates should be under change control before deployment, including “any change to physical objects the model is using as input”. The draft also asks for configuration control with effective measures to detect unauthorised change. It also asks for the model’s performance metrics to be monitored regularly to catch changes in the computerised system, naming deterioration or change of a lighting condition as an example.

We recommend recording the baseline as a single versioned record linked to the PQ report. It would contain:

• the algorithm or model version, with a reference to the training data and acceptance criteria it was qualified against
• every decision threshold and reject parameter, with the tolerance band (if any) that PQ actually covered
• camera and lens identity plus acquisition settings such as exposure, gain and focus
• lighting type, geometry and measured intensity at the inspection point
• container handling settings such as line speed and spin, since handling changes are named alongside optics and software as requalification triggers
• the identity of the qualification set and function set, and the detection limits set at PQ

Once that record exists, the question “does production still match what was qualified?” stops being a matter of belief. It becomes a check you can run and audit, comparing live version hashes, threshold values and lux readings against the record.

Draft Annex 22 also asks for detection of two risks that no change request announces. The first is unauthorised change, such as informal threshold tweaks. Each one may look small, but together they move the machine away from the conditions PQ proved. The second is deterioration of the lighting condition, such as lamp wear. The ECA’s 2014 good practice paper on visual inspection treats illumination as something to qualify regularly within the maintenance programme, suggesting a six-month check interval. Spotting quiet optical degradation across a changing product mix is a monitoring problem in its own right. Here we only need it as a reason the baseline must include measured lighting values, not just part numbers.

The change control process for a vision inspection machine after its PQ run

We recommend putting every proposed change to the inspection machine through the same ordered path after PQ.

1. Describe and justify. FDA’s 2011 process validation guidance, citing 21 CFR 211.100, says a description of the planned change, a well-justified rationale, an implementation plan and quality unit approval must be documented before implementation.
2. Classify the change. PIC/S PI 054-1 (2021) notes that change procedures often assign a risk-based classification such as critical, major or minor, together with an impact assessment.
3. Evaluate with the right people. ICH Q10 calls for expert teams from the relevant areas to evaluate proposed changes, with evaluation criteria set in advance. For a vision system, that team should include whoever understands the model as well as quality and production.
4. Decide the test scope. Under draft Annex 22, a change is evaluated to decide whether the model needs retesting, and any decision not to retest must be fully justified.
5. Approve. Annex 15 requires changes to be authorised and approved by the responsible persons under the pharmaceutical quality system.
6. Implement and verify. Demonstrate effective implementation against the acceptance and effectiveness criteria predefined in change planning.
7. Evaluate the first batches and the change’s effectiveness. ICH Q7 section 13.15 asks for an evaluation of the first batches produced or tested under the change. Annex 15 section 11.7 asks for an effectiveness evaluation where appropriate.
8. Update the baseline record. A successfully implemented and verified change can become part of a new validated baseline. That only holds if the record is updated to show it.

Between changes, two separate controls keep the machine honest. Function sets are abridged kits of fairly crude defects run before and after each batch to show the system still works. ECA guidance says routine results should stay within the detection limits defined at PQ. Function tests are not requalification, though. ECA Academy points out that they focus on sensor functionality and, because their scope is limited, cannot replace requalification.

For periodic requalification, ECA Academy recommends an AVI system be requalified ideally every year and at least every two years. That requalification must include an evaluation of changes and deviations during the period and a statistical trend analysis of routine performance data. Confiancapharmazon suggests a tighter 6 to 12 month interval, or at least annually, and after any equipment modification. Either way, Annex 15 section 4.2 requires the chosen period to be justified, its evaluation criteria defined and the possibility of small changes over time assessed.

Change control cannot show changes that were never recorded, which is why Annex 11 asks for consideration, based on a risk assessment, of a system-generated audit trail of GMP-relevant changes, with reasons documented and the trail reviewed regularly.

Which changes should trigger requalification of a vision inspection system?

Treat any change that touches image formation or decision logic as presumed to need requalification until an impact assessment shows otherwise. Confiancapharmazon’s guidance on AVI qualification kits is blunter: any change in optics, software or handling systems requires requalification. PIC/S lists equipment upgrades, including computerised systems, among the common triggers for change.

Other sources allow a lighter response in some cases. One validation guide notes that replacing a component with an equivalent part that doesn’t alter functionality may need only limited verification. We read these as compatible rather than conflicting. Like-for-like is a claim the impact assessment has to prove for the vision chain, not something it can assume.

Requalification decision table for common changes

Change	Why detection may shift	Typical response
Algorithm retrain or new model version	The decision logic itself changes	Retest against the qualification set and PQ acceptance criteria. Any decision not to retest must be fully justified
Threshold moved beyond the range qualified at PQ	Reject boundaries move outside the evidence base	Requalification before production use
Threshold adjusted within a range PQ explicitly covered	Still inside the qualified evidence	Documented rationale plus a function-set check
Lamp or LED replaced with an equivalent part	Illumination is a qualified condition that needs regular checking	Measure the illumination, run the function set, record the rationale. Escalate if results fall outside PQ limits
Lighting type or geometry changed	Defect contrast changes	Requalification
Camera or lens replaced	Resolution, focus and sensor response change	Requalification
Handling change (line speed, spin, container path)	Motion and presentation of defects change	Requalification
Software change outside the detection path (reporting, UI)	Annex 11 still requires computerised-system changes to follow a defined procedure	Regression testing and documentation for the computerised system

Requalification of the hardware itself is covered in our piece on change control and requalification for vision inspection hardware.

Running the impact assessment: documentation only, partial retest or full requalification

The impact assessment decides how much evidence a change needs, and the response should be proportionate rather than automatic. One validation guide lists possible responses running from a documentation update, through targeted functional testing and partial requalification, to full revalidation. Its goal is enough evidence that the change works without invalidating existing controls, not a mechanical repeat of the original validation. PIC/S says the formality, effort and documentation should match the level of risk. For simple low-risk changes, an impact assessment alone can be enough to document the rationale. More complex or significant changes call for a more formal risk assessment.

ICH Q7 section 13 supports the same approach. It says a classification procedure, for example minor or major, helps decide the level of testing, validation and documentation, and that scientific judgement determines what extra studies are appropriate. Annex 15 asks the assessment to cover product quality, documentation, validation, calibration, maintenance and other systems, so that verification or requalification can be planned.

For a vision machine, one question to ask is whether the existing qualification evidence still represents the changed condition. If the PQ challenge runs were done with the old lens, they say nothing about the new one, however good its specification sheet looks. If the threshold moved within a band PQ already tested at both ends, the evidence still holds. A second check is worth making: equipment qualification shows the equipment works, but it doesn’t necessarily show the changed process still produces equivalent product.

In practice the outcomes fall into three bands:

• Documentation only, when the change is simple and low risk and the impact assessment records the rationale.
• Partial retest, such as targeted functional testing, verification or regression testing of the affected software, proportionate to the change and the strength of existing evidence.
• Requalification against the PQ acceptance criteria, when optics, software or handling has changed.

Whichever band applies, the acceptance and effectiveness criteria should be set before implementation, not after the results arrive.

What FDA and ICH guidelines expect from change control for automated inspection equipment

Change control for automated visual inspection draws on several documents, each covering a different layer:

• FDA, 21 CFR 211.100 (as cited in FDA’s 2011 process validation guidance): documented description, rationale, implementation plan and quality unit approval before a change is implemented. The same guidance cites 21 CFR 211.68, which requires automated, mechanical and electronic equipment to be calibrated, inspected or checked under a written programme. It also treats a system for detecting unplanned departures from the validated process as essential in continued process verification.
• ICH Q7, section 13: a formal change control system to evaluate all changes that may affect production and control, with written procedures covering equipment, including computer hardware.
• ICH Q10: an effective change management system to evaluate, approve and implement changes, with the effort of evaluation matched to risk and a post-implementation check that the objectives were met without harm to product quality.
• EU GMP Annex 15: planned changes that may affect quality are documented and their impact on validated status assessed, with justified requalification periods under section 4.
• EU GMP Annex 11: changes to a computerised system, including its configuration, only under a defined procedure, and periodic evaluation covering deviations, incidents and upgrade history.
• Draft EU GMP Annex 22 (2025 consultation, not yet final): change control over model, system and process, configuration control, and performance monitoring that would catch lighting deterioration.

In its July 2023 warning letter to Baxter Healthcare, FDA said automated methods should be rigorously studied for capability and robustness under various conditions, machine settings, container-closure sizes and defect types. The letter also criticised the lack of scientific justification for removing harder-to-detect particles from the defect kits used for qualification. Whether a defect library is representative enough to be defensible is its own question. For change control, the point is that machine settings fall within the scope of the robustness FDA expects you to show.

CAPA or change control when an inspection system’s performance drifts?

CAPA addresses the cause of a problem, and change control governs implementing the fix. ICH Q10 defines corrective action as action to eliminate the cause of a detected nonconformity, and preventive action as action to prevent a potential one from occurring. A deviation shows something unexpected happened, the investigation finds the cause, CAPA addresses that cause, and change control governs implementing any modification the CAPA requires. Validation then decides what evidence the modified condition needs.

ICH Q10 treats CAPA and monitoring outputs as drivers of change that the change management system then evaluates, approves and implements. One validation guide warns that CAPA should not become an informal substitute for change control just because the modification is corrective. The CAPA explains why action is needed. Change control explains how the modification is assessed, approved, implemented and verified.

Take an AVI machine whose function-set results trend downwards over several weeks. If critical defects are being missed, a CAPA should be opened. Say the investigation finds an ageing illumination source. Replacing it is still a change to the qualified baseline, so it goes through impact assessment and the requalification scope set out above. The documented path runs: signal, investigation, cause, CAPA, controlled change, validation impact, implementation, effectiveness. Closing the CAPA should confirm that the change controls are complete, any requalification is done, and the revised validated baseline is documented.

Where the engineering ends and the quality decision begins

We see change control on a qualified inspection system as release-gate work: the part of production AI reliability that stops a silent configuration change from reaching product, which is a separate matter from how accurate the model was on the day of PQ. TechnoLynx’s AI engineering for life sciences covers the baseline definition, the logging, the drift detection, the trigger logic and periodic robustness checks of the algorithm against lighting and fixture variation. Approval of each change stays with the responsible persons under your quality system. Teams with strong computerised-system validation staff and stable, vendor-locked inspection recipes may need only better procedures, not engineering help.

Start by answering one question for your own line: if an inspector asked today whether the algorithm, thresholds and lighting in production match the PQ record, could you show it from a log in minutes, or would you have to reconstruct it from memory and maintenance tickets?