In drug manufacturing, aseptic manufacturing is how sterile drugs are made when terminal sterilisation of the product is not feasible. The drug, the container and the closure are each sterilised separately, then filled and sealed in an extremely clean environment. Nothing sterilises the filled unit afterwards, so sterility depends on the critical zone, where human interventions pose the biggest risk. Media fills and microbial sampling check that risk periodically and indirectly. ML-analysed video monitors every operator intervention in real time and documents it objectively. It adds to the sterile controls an aseptic line depends on rather than replacing them.
Aseptic manufacturing under FDA guidance and EU Annex 1
FDA’s position is that sterile drugs should be made by aseptic processing only when terminal sterilisation is not feasible. The reason is plain: FDA’s sterile drug inspection programme states that aseptic processing carries a higher risk of microbial contamination than terminal sterilisation, and that any manual or mechanical manipulation of sterilised drug, containers or closures before or during filling poses that risk.
In the US, the rule base is 21 CFR 211.113(b): written procedures designed to prevent microbiological contamination of drugs purporting to be sterile must be established and followed, including validation of any sterilisation process. FDA’s 2004 guidance on aseptic processing sets out the agency’s current thinking on meeting that rule. Its compliance programme treats firms that follow the guidance as generally compliant, while allowing other approaches that satisfy 21 CFR Parts 210 and 211.
Annex 1 defines Grade A as the critical zone for high-risk operations such as the filling zone, the stopper bowl and aseptic connections under first air. It requires a Contamination Control Strategy (CCS) that defines all critical control points and assesses how effective every design, procedural, technical and organisational control and monitoring measure is. The CCS also covers an assessment of the feasibility of scientifically sound alternative methods that optimise detection of environmental contamination. That clause is where the CCS assesses whether a new monitoring method is feasible.
A sterile line also needs a lot more than monitoring: Class 100 (ISO 5) air at the work site, HEPA integrity and velocity checks, validated filtration and integrity testing, and container sterilisation and depyrogenation. Annex 1 says outright that monitoring or testing alone does not give assurance of sterility. That sentence sets the limit for everything that follows.
Why media fills and sampling give limited assurance
The media fill, which Annex 1 calls aseptic process simulation, is how an aseptic processing operation is validated. Growth medium replaces the product, runs through the same equipment, environments and manipulations, and the sealed units are incubated to detect contamination. FDA recommends at least three consecutive successful runs for initial line qualification, followed by semi-annual requalification of each line. FDA’s guidance gives 5,000 to 10,000 units as a generally acceptable starting run size. Below 5,000, the run should at least equal the maximum batch size.
For runs of 5,000 to 10,000 units, one contaminated unit triggers an investigation and two are cause for revalidation. FDA itself acknowledges scientific and technical limits on how precisely process simulations can characterise a system of controls.
The authors of an ISPE case study set out the weakness as a chain. For a media fill vial to show growth, a microbe must be on the glove, the operator must enter the critical zone, the microbe must transfer to the zone, enter a container and then grow during incubation. The authors treat this sequence as one event and say media fills may have an inherently low probability of detection, which makes them best at catching gross microbial problems. A media fill also does not always reproduce production faithfully.
Routine environmental sampling has its own blind spots. FDA notes that environmental monitoring methods do not always recover the microorganisms present, and that low-level contamination is particularly hard to detect. FDA’s inspection programme adds that taking more sterility test samples does not greatly improve the chance of detecting contamination present at a very low level in a lot.
How does AI-assisted aseptic monitoring differ from traditional sterile manufacturing controls, and where does the vision-plus-signal approach add value over conventional sterility assurance?
Traditional controls measure contamination indicators, such as colony counts on plates, growth in media fill units and particle counts. AI-assisted monitoring detects the risky event itself, an operator intruding into the critical zone. The ISPE authors point out that there is currently no objective determination of whether an operator intruded into the critical zone during an intervention. Operators and secondary observers catch obvious cases, but not every instance. Their approach uses video and a machine learning model that analyses every intervention in real time and gives 100% automated monitoring of operator interventions.
Their risk assessment says the criticality of the data does not change. What changes is detectability: human visual monitoring rates moderate to moderate-low because borderline cases are missed, while video with ML analysis gives a high probability of detection. They also describe the ML data as more accurate, contemporaneous, consistent and complete than human observation during media fills and production.
By “vision-plus-signal” we mean pairing that camera stream with the line’s existing environmental signals. FDA inspectors check whether pressure differentials, temperature and humidity are continuously monitored and alarmed during production. Annex 1 requires critical pressure differences to be continuously monitored and recorded, with a warning system. In our view, the value comes from putting the two on one timeline. An intrusion flag next to a particle excursion or a pressure dip is far easier to investigate than either alone, and the combined record lets the quality unit reconstruct an intervention rather than infer it.
Some care with the ISPE claim is needed. The authors argue that ML monitoring may remove the need for media fill assessment of operator interventions in isolator or RABS suites, provided the full control strategy is shown to mitigate all sterility failure risks. That is the authors’ argument from their own case. Our reading is that it strengthens the CCS and the investigation record, and media fills stay in the programme. Our own work on the topic takes the same line: AI-assisted line monitoring is not meant to replace sterile filters, steam sterilisation or other sterile controls, and manual visual inspection rarely produces an audit-grade record. How that works day to day on a line is covered in our piece on AI-driven aseptic operations and contamination reduction.
Where on an aseptic line does AI monitoring cut contamination risk the most?
Human interventions into the critical zone pose the biggest risk to product sterility. FDA’s inspection programme names humans as a significant source of contamination, especially on lines where operators routinely enter the critical area. FDA’s guidance says that as operator activity increases, the risk to finished product sterility increases. The ISPE authors call operator interventions the current “weak link” in the aseptic control strategy.
FDA tells its investigators to watch these operations in real time:
- line set-up, especially lines needing multiple aseptic assemblies or lacking SIP of the product pathway;
- handling of equipment jams and stoppages;
- personnel and material flow in the aseptic filling room;
- the nature and frequency of interventions during filling.
Those are the moments a camera should be watching. Two more deserve attention: transfer of partially stoppered vials to a lyophiliser, which Annex 1 requires under Grade A conditions with minimal direct operator intervention, and open-door RABS interventions. FDA’s Rick Friedman, speaking at an ISPE event, described firms running routine open-door RABS interventions without documented justification or duration limits.
The benefit depends on the line’s design. Isolators completely separate the filling line from the external environment, while RABS and blow-fill-seal reduce human interventions in the critical area. So in our judgement, intrusion detection pays off most on conventional lines and open RABS, where interventions are frequent and doors open. On isolators and RABS, Friedman’s account also flags mechanical failure modes: glove integrity, gasket wear at transfer ports and loss of pressurisation that building management systems must detect at once. There, the signal half of vision-plus-signal does more of the work than the camera. If the decision is about the barrier or automation itself rather than about monitoring, that belongs with fill-finish solutions for pharma manufacturing.
What evidence does an AI-assisted fill-finish system need for Annex 1 inspectors?
Three expectations bear on that evidence: equivalence or superiority to established methods, operation inside the model’s sample space, and interventions documented for Quality Unit review. Annex 1 section 9.28, quoted in the ISPE study, allows rapid and automated microbial monitoring methods once validation has shown they are equivalent or superior to established methods. For automated inspection of filled containers, Annex 1 paragraph 8.32 requires validation to detect known defects and performance equal to or better than manual inspection.
Draft EU GMP Annex 22, a consultation text, applies to machine learning models used in critical GMP applications with a direct impact on patient safety, product quality or data integrity. It asks for a detailed intended use, including a full characterisation of the input sample space with common and rare variations, approved by a process subject matter expert before acceptance testing. ISPE’s GAMP Guide: Artificial Intelligence builds on the risk-based approach of GAMP 5 (Second Edition) and covers monitoring strategy and ongoing control during operation.
Evidence pack an inspector can review
| Evidence | What it demonstrates | Where the expectation comes from |
|---|---|---|
| Intended use and input sample space, SME-approved | The model is scoped to this line, these interventions and these conditions | Draft Annex 22 |
| Test metrics suited to the intended use (e.g. confusion matrix, sensitivity, specificity, accuracy, precision, F1) | Detection performance is measured, not asserted | Draft Annex 22 |
| Equivalence or superiority study against human observation | The method meets the bar for alternative monitoring methods | Annex 1 9.28 as cited by ISPE |
| Live process validation in the actual aseptic process | Performance holds outside development data | ISPE case study |
| Feature attribution or heat maps on test-data decisions | Reviewers can see which features drove a classification during testing | Draft Annex 22 |
| Documented interventions reviewable by the Quality Unit | Interventions into ISO 5 are recorded and evaluated | FDA compliance programme |
| Ongoing performance and input-drift metrics | The model is still operating inside its validated space | Draft Annex 22 |
| CCS entry with risk assessment and residual risk | The method’s role in contamination control is defined | Annex 1 |
TechnoLynx treats this kind of evidence as an engineering output: artefacts a quality owner or auditor can read, defend and sign against, built from the system itself rather than written as policy. That boundary is real. Building the artefacts is not regulatory sign-off, certification or a promise that an inspection will go well.
How is line monitoring instrumented without disrupting qualified processes?
In our view, intervention monitoring stays a manageable change if it follows three engineering choices:
- Freeze the model. Lock the weights and release them through change control, never as a model that learns in production.
- Keep the hardware out of the process boundary. Cameras and compute sit outside Grade A and take no part in product contact or in the line’s control loop.
- Read existing signals, don’t add actuators. Environmental data is consumed from the monitoring systems already qualified, so the monitor observes and records without changing any process parameter.
Keeping the existing qualification strategy is a goal, not a given. A PDA paper on machine learning for automated visual inspection of filled injectables says ML implementation does not require changes to current AVI qualification strategies, and that freezing the trained model and putting it through current validation gives good assurance of reliable production performance. That paper concerns container inspection; carrying it over to intervention monitoring is our judgement. The ISPE case study makes a similar argument: the ML-analysed data is itself the evidence for validating the change, because it shows more control and lower process risk.
FDA recommends evaluating air samplers for suitability in an aseptic environment, including disruption of unidirectional airflow. Annex 1 expects airflow visualisation to show no ingress from lower to higher grade areas and no air travelling over operators or equipment that may transfer contamination to higher grade areas. A camera bracket near first air should be checked against the smoke studies for the same reasons, which is our judgement rather than a written rule. Annex 1 also expects equipment monitoring requirements to be set in user requirement specifications and confirmed during qualification.
Where do aseptic AI deployments fail in production?
The first failure mode is drift in the scene. Draft Annex 22 asks for model performance to be monitored regularly to detect changes such as deterioration of, or a change in, a lighting condition, and for metrics that detect drift in input data away from the model’s sample space. The prevention is to put the camera environment under change control together with the model and to alarm on input drift, not only on missed detections.
The second is training data that does not represent production. FDA’s 2023 discussion paper on AI in drug manufacturing notes that the robustness of the underlying data shapes the model’s decisions and makes unintended bias harder to avoid. Rare interventions, such as a jam on the last needle, need to be in the intended use and the test set, as draft Annex 22 asks for rare variations.
The third is a model that keeps learning. FDA’s paper says continuously learning systems may challenge regulatory oversight, because the current approach updates models through change control in the pharmaceutical quality system. The fix is the frozen model described above.
The fourth is a flag nobody reviews. Where a model feeds a human decision and testing effort has been reduced, draft Annex 22 expects records of the review process, possibly a review of every output. Annex 1 expects alarm events to be acknowledged, trended and assessed at a frequency matched to criticality. A monitor whose flags pile up unread fails the CCS and produces nothing an inspector can use. FDA’s Friedman, citing ICH Q9(R1), warned against treating any new technology as a cure-all before fitness for intended use is established.
Which fill-finish KPIs become contractable with vision-plus-signal monitoring?
No measured source supports specific KPI targets here, so this section is judgement. A KPI becomes contractable, meaning usable in a quality agreement or a CDMO contract, when it is measured objectively, completely and with an audit trail. Before continuous monitoring, intervention records rested on subjective human observation. Afterwards they can be counted.
Candidates we would consider:
- interventions into Grade A per batch, by type (planned versus corrective);
- confirmed critical-zone intrusions per batch and per shift;
- intervention duration, and open-door time on RABS lines;
- time from intrusion flag to Quality Unit review;
- acknowledgement time for critical environmental alarms;
- share of interventions with a complete, linked video and signal record.
Two cautions apply. Sterility itself is not one of these KPIs: Annex 1’s statement that monitoring alone gives no sterility assurance applies to contract metrics too. Also, a KPI attached to payment changes behaviour, so a target on intrusion counts should never discourage an operator from making a necessary intervention. Annex 1 already expects container inspection results to be recorded and defect and reject levels trended statistically.
Does continuous manufacturing change the picture for aseptic fill-finish monitoring?
ICH Q13 covers development, implementation, operation and lifecycle management of continuous manufacturing for drug substances and drug products, including converting existing batch products, and became legally effective in the EU on 10 July 2023. FDA’s inspection programme expects at least one semi-annual media fill per line per shift, except on isolator lines. Residence time distribution, process models and the wider AI model lifecycle for continuous lines are covered in our article on continuous manufacturing and AI process control.
For the broader context of AI across clinical and life sciences settings, our overview of artificial intelligence in healthcare places manufacturing alongside the other regulated uses. What it costs when a sterile batch is lost is examined in pharmaceutical batch failure and how AI prevents it.
The decision still sits with the CCS
The real question for an aseptic site is whether intervention monitoring has a defined place in its Contamination Control Strategy, with acceptance criteria, effectiveness review and documented residual risk, as Annex 1 asks of aseptic process controls. A model that detects intrusions accurately but sits outside the CCS is a camera with good statistics. One written into the CCS turns the weakest, least visible part of aseptic manufacturing into something the Quality Unit can review batch by batch.
FAQ
How is aseptic technique different from sterility?
For sterility, FDA expects a sterilisation process to demonstrate a sterility assurance level of 10⁻⁶ or better. Aseptic technique is the operator practice that keeps already-sterilised items sterile while they are handled, and FDA says operators in aseptic activities must use it at all times.
