Remplacer les silos de conformit� par un syst�me de contr�le vivant

Un lot est bloqu� parce qu'une d�rive de proc�d� n'est d�tect�e qu'au contr�le qualit� final, des heures apr�s le pr�l�vement qui l'aurait r�v�l�e � temps. Une rupture d'approvisionnement atteint un service de soins avant que l'alerte interne n'ait �t� prioris�e parmi les dizaines d'autres remont�es de la semaine. Un signal de pharmacovigilance est class� sans investigation, noy� dans un flux qu'aucune m�thode de consensus ne filtrait. Trois sc�narios distincts, un m�canisme unique : le signal existait, mais n'a pas �t� synth�tis� � temps pour d�clencher une d�cision.

L'hypoth�se de cette synth�se est la suivante : une organisation pharmaceutique peut remplacer les v�rifications de conformit� isol�es par un syst�me de contr�le vivant lorsque les donn�es de proc�d�, de produit et de qualit� sont align�es sur un contexte de lot/shared commun ; �valu�es conjointement avec des mod�les multivari�s robustes et adaptatifs ; connect�es � des actions de contr�le pr�d�finies ; et conserv�es comme preuve tout au long du cycle de vie sous gouvernance humaine et mod�le.

Article disponible en anglais uniquement. La version compl�te de cet article est publi�e en anglais. Pour lire la synth�se d'evidence compl�te, veuillez passer en anglais.

A batch is held because a process deviation is detected only at final quality control, hours after the sample that would have revealed it in time. A supply disruption reaches a care ward before the internal alert has been prioritised among the dozens of other escalations from the week. A pharmacovigilance signal is closed without investigation, buried in a flow that no consensus method filters. Three distinct scenarios, one mechanism: the signal existed, but was not synthesised in time to trigger a decision.

The hypothesis of this synthesis is that a pharmaceutical organisation can replace isolated compliance checks with a living control system when process, product and quality data are aligned to a shared batch/run context; evaluated jointly with robust multivariate and adaptive models; connected to predefined control actions; and retained as lifecycle evidence under human and model governance.

Part 1: Build one evidence plane without erasing operational boundaries

Compliance islands arise when instruments, unit operations, quality databases and reviews each retain their own identifiers, time bases, definitions and thresholds. The first implementation problem is therefore data and control integration, not visual aggregation alone.

Robert C. Martin's Clean Architecture supplies a useful structural analogy. In Chapter 22, the Dependency Rule states that dependencies should point toward higher-level policy. Applied cautiously to pharmaceutical control, source systems and instruments remain replaceable mechanisms; the control strategy, product knowledge and decision rules are the higher-level policy. Integration should therefore standardise interfaces and evidence objects without allowing one vendor format or local database to become the control strategy itself.

Kim et al. (2021) reviewed PAT tools across pharmaceutical unit operations and positioned PAT as a control-strategy component for continuous process verification. Casian et al. (2022) reviewed data fusion in PAT and identified the integration chain that must be solved: probe and sampling integration, heterogeneous data collection, model construction, connection to control systems, calibration and validation of the integrated system.

ICH Q10 requires a process performance and product quality monitoring system capable of assuring continued process capability and identifying opportunities for improvement. Its 2024 Q&A update states that increasingly complex information from QbD, PAT, real-time generation and control-monitoring systems must be captured, managed and shared across the product lifecycle. FDA lifecycle validation guidance similarly requires an ongoing programme that collects, statistically trends and reviews process and product data.

The Pretzner case study operationalises these principles. Data from separate formulation, filtration, lyophilisation and quality-database sources were aligned to the lot, separated into process phases and reduced to a common feature matrix before robust PCA. The empirical plot is therefore not evidence that centralisation alone creates control; it is evidence that integration and contextual alignment make system-level multivariate monitoring possible.

Part 2: Detect joint drift and preserve an explainable path back to the process

Once data are integrated, the control system must distinguish ordinary correlated variation from material multivariate deviation and must expose the variables and process intervals that drove the alert. A black-box score without localisation merely creates a new compliance island.

Pretzner et al. (2020) provide the strongest real-manufacturing evidence for this design rule. Their study combines real formulation, fill, and finish data with time-series alignment, phase-aware feature extraction and robust principal component analysis. A model built from 58 lots uses 21 principal components and explains 80.88% of variance. When the robust PCA model is applied, the most prominent score-distance outlier is clearly identified, while lots near the model centre are correctly classified as normal.

The contrast with classical PCA is instructive. On the same 58-lot feature matrix, classical PCA flags 4 score outliers and 9 orthogonal outliers. Robust PCA flags 1 score outlier and no orthogonal outliers, with 7 and 1 moderate cases respectively. The authors could not identify a process reason for most classical-PCA flags and attribute them to the nonrobust model. This shows that detector choice materially changes the alarm burden.

The contribution analysis demonstrates the transition from "this batch differs" to "these variables, in this phase, account for the difference." For the outlying lot 22, contributions from 252 variables show that QDB-1 and phase-derived LP1/LP2 features dominate. The corresponding phase-trajectory plot reveals that lot 22 has approximately 30% higher noise than the central lot 54, linked to the robust-PCA score distance. This converts a multivariate alarm into a specific, phase-localised investigation target.

Part 3: Connect monitoring to adaptive models, predefined actions and governed learning

A living system must continue to function as materials, equipment, process conditions and data distributions change. Static models and isolated alarms cannot satisfy this requirement. Adaptation must be controlled: new data can update the monitoring model only through versioned rules, performance checks and accountable approval.

Chen et al. (2020) reviewed digital twins in pharmaceutical and biopharmaceutical manufacturing and identified the combination of PAT, process modelling and data integration as the technical base for a model that remains connected to the physical process. The review does not establish a universal production architecture, but it shows why a "living" control system requires an executable model and continuous data exchange rather than a static report.

Khuat et al. (2025) provide recent empirical evidence for the model-lifecycle problem. In their case studies, models that were not updated performed poorly under conditions that differed from the original training data, while real-time update, just-in-time learning and mixture-of-experts approaches improved tracking. The manuscript demonstrates industrial potential, but it remains a limited set of cell-culture case studies; deployment would require independent validation, version control, alert-performance qualification, calibration maintenance and predefined human intervention.

Thakur et al. (2024) provide a complementary peer-reviewed systems case. Their cyber-physical production system operated an integrated continuous monoclonal-antibody manufacturing train for 55 hours at steady state. In-line and at-line sensors supported end-to-end data collection; deviations in chromatography and ultrafiltration were monitored with multivariate analysis and in-process controls, while surge tanks supported corrective action and temporary decoupling.

ICH Q13 supplies the regulatory architecture for this operational link. Its examples combine PAT measurements with predefined limits, material diversion, operator investigation, process stop and surge capacity. It also describes real-time monitoring, trending and prediction based on NIR, feeder and process-parameter data. FDA continued process verification requires systems capable of detecting unplanned departures and ongoing statistical analysis by trained personnel. These sources converge on a controlled action loop; none support autonomous model-driven release without the approved control strategy and accountable quality oversight.

Conclusion

The evidence supports a narrower and defensible proposition:

Replacing compliance islands requires a governed control architecture that joins distributed evidence through shared context, evaluates correlated variation with robust and adaptable models, localises the drivers of abnormality, and binds monitoring outputs to predefined actions and lifecycle evidence.

The evidence does not establish that any single platform, digital twin, PAT tool or machine-learning model guarantees a state of control. The decisive capability is the integrated operating system around those components: data integrity, semantic alignment, model validity, control-strategy linkage, accountable intervention and retained decision lineage.

Industrial benchmark capability stack

Missing information and evidence gaps

• No prospective, multi-site pharmaceutical study was located that compares an integrated living control system with siloed compliance practice on batch rejection, deviation recurrence, review time or shortage outcomes.
• No formal meta-analysis was located for the exact architecture question.
• The adaptive-ML evidence is recent and case-specific; external validation and final peer-review status must be checked before regulated deployment.
• Reported model metrics cannot be transferred between products, modalities or sites without a representative reference population and predefined acceptance criteria.
• The literature does not define a universal minimum data architecture, model class or automation level for all pharmaceutical processes.

Sources

• Casian, T., Nagy, B., Kovacs, B., Marosi, G., & Nagy, Z. K. (2022). Challenges and opportunities of implementing data fusion in process analytical technology - A review. Molecules, 27(15), 4846. doi.org/10.3390/molecules27154846
• Chen, Y., Yang, O., Sampat, C., Bhalode, P., Ramachandran, R., & Ierapetritou, M. (2020). Digital twins in pharmaceutical and biopharmaceutical manufacturing: A literature review. Processes, 8(9), 1088. doi.org/10.3390/pr8091088
• Food and Drug Administration. (2011). Process validation: General principles and practices. Stage 3 - Continued Process Verification, pp. 16-17.
• International Council for Harmonisation. (2008). ICH Q10 Pharmaceutical Quality System. Sections 3.1.3 and 3.2.1, pp. 10-11.
• International Council for Harmonisation. (2022). ICH Q13 Continuous Manufacturing of Drug Substances and Drug Products. Main guideline and Annexes, especially pp. 23, 27-30.
• International Council for Harmonisation. (2024). Q8/Q9/Q10 Questions & Answers (R5). Knowledge Management, pp. 13-14.
• International Council for Harmonisation. (2025). Reflection paper on advanced manufacturing technologies. Endorsed 8 October 2025.
• Kim, E. J., Kim, J. H., Kim, M. S., Jeong, S. H., & Choi, D. H. (2021). Process analytical technology tools for monitoring pharmaceutical unit operations: A control strategy for continuous process verification. Pharmaceutics, 13(6), 919. doi.org/10.3390/pharmaceutics13060919
• Khuat, T. T., Peng, J., Bassett, R., Otte, E., & Gabrys, B. (2025). Lessons learned from deploying adaptive machine learning agents with limited data for real-time cell culture process monitoring. arXiv. doi.org/10.48550/arXiv.2509.02606
• Martin, R. C. (2017). Clean architecture: A craftsman's guide to software structure and design. Pearson. Chapter 22, pp. 201-209.
• Pretzner, B., Taylor, C., Dorozinski, F., Dekner, M., Liebminger, A., & Herwig, C. (2020). Multivariate monitoring workflow for formulation, fill and finish processes. Bioengineering, 7(2), 50. doi.org/10.3390/bioengineering7020050
• Thakur, G., Saxena, N., Yezhuvath, V. B., & Rathore, A. S. (2024). A cyber-physical production system for the integrated operation and monitoring of a continuous manufacturing train for the production of monoclonal antibodies. Bioengineering, 11(6), 610. doi.org/10.3390/bioengineering11060610

Evidence cut-off: 3 August 2026. Source figures are reproduced from open-access publications with attribution. Long captions are paraphrased; figure number, page/section, and source DOI preserve traceability.