专家分析 AI 与机器学习在质量回顾中的应用:从孤立 KPI 到关联信号
AI and Machine Learning in Quality Reviews: Turning Data into Signals
GMP Journal 刊出基于 ECA 课程的分析文章,讨论如何用可解释的 AI 和机器学习把质量与运营数据关联,支持质量回顾中的信号识别。文中匿名示例把偏差逾期与运营延误小时合并观察,指出结构化 QA 与生产沟通后两项指标同时下降,但作者明确该示例不支持因果结论。
21.10.2026
GMP/GDP – On Demand Online Training
You can book the desired online training from our extensive database at any time. Click below for more information.
Stay informed with the GMP Newsletters from ECA
The ECA offers various free of charge GMP newsletters for which you can subscribe to according to your needs.
The article is based on the session "AI and ML Insights for Quality Reviews" at the ECA Course "Improve your Quality Reviews" (June 2026), delivered by Simone Baisi, Global GMP & GDP Quality Compliance Specialist and Microsoft Copilot Champion, Kedrion Biopharma*, Italy.
GMP organisations collect large amounts of quality data, but much of it is still analysed in isolation. As a result, Product Quality Reviews (PQRs), Annual Product Reviews (APRs) and Management Reviews often describe what happened without explaining why it happened or what it means for operations. This article shows how explainable AI and Machine Learning (ML) can help QA connect quality and operational data, detect risk signals earlier and improve decision-making, without replacing GMP judgement.
The perspective is deliberately practical: the value of AI in Quality Reviews is not the model itself, but the ability to ask better management questions based on connected data.
From Isolated KPIs to Connected Signals
In a typical GxP environment, QA teams work with deviations, CAPAs, complaints, OOS/OOT results, analytical trends, audit findings, supplier performance and training data. The issue of inefficient quality reviews is rarely a lack of data; the issue is how these data are reviewed. From a Quality System and Compliance perspective, this is where the real opportunity sits: not simply producing more KPI tables, but turning existing indicators into a more connected view of quality performance and risk.
Very often, each KPI lives in its own section of the review: deviations here, CAPAs there, complaints somewhere else, and operational performance in a different report. The real signal, however, frequently sits between indicators - for example, between overdue deviations and production delays, or between CAPA effectiveness and recurrent issues.
Traditional reviews therefore tend to be descriptive: deviations increased, CAPA closure was delayed, complaints remained stable. This is necessary, but not sufficient. In practice, the management question should move from "What happened?" to "What is this combination of signals telling us, and where should management attention be focused?"
A useful quality review should also explain relationships: why did a trend change, what else changed at the same time, and did a quality trend create operational risk? This is where AI and Machine Learning can support QA - not by deciding, but by highlighting patterns and connections that are difficult to see in isolated KPIs.
A Practical Case: Connecting Overdue Deviations with Operational Impact
The case study is an illustrative and anonymised example intended to explain the analytical approach rather than disclose company-specific performance data.
In the case study, one production area showed two worrying trends: increasing overdue deviations and increasing operational delay hours. At first glance, this looked like a simple backlog problem. A closer review suggested something more important: late QA involvement was creating documentation pressure; Production had to reconstruct information long after the event; key resources were pulled into follow-up work instead of supporting execution; and the quality backlog became an operational constraint.
To understand this, quality and operational data were combined in one view. On the quality side, overdue status, deviation aging, investigation cycle time, QA follow-up frequency and missing documentation were considered. On the operational side, delay hours, planning impact, batch rescheduling and resource pressure were included. The value came from reviewing these data together and asking: “What story do they tell when viewed as one picture?”
The intervention was deliberately simple: structured QA-Production meetings to review open deviations early, clarify documentation needs and agree on priorities. QA moved from reactive chasing (“please send the missing information”) to proactive support. After implementation, overdue deviations and operational delays decreased and documentation pressure eased. The main insight was that more structured deviation management was associated with improved operational performance, and that this relationship became visible only when quality and operational indicators were reviewed together (see Figure 1).

Figure 1: Monthly trend in overdue deviations and operational delay hours before and after structured QA-Production support.
Illustrative and anonymised data. The two series use separate y-axes because they have different units. The dashed vertical line indicates the intervention point between July and August. The figure supports signal detection and temporal comparison; it does not demonstrate causality.

Figure 2: Relationship between overdue deviations and operational delay hours by operating phase.
Illustrative and anonymised scatter plot. Each point represents one month; colours identify the predefined reactive, transition and structured-support phases, while the X markers show the phase centroids. The plot supports behavioural grouping and pattern recognition; it should not be interpreted as the output of an unsupervised clustering algorithm, as causal proof or as automated classification.
How Explainable AI and ML Support Quality Reviews
The session focused on practical, explainable methods rather than complex black-box models. In a GxP context, a simpler model that can be understood, challenged and documented is often more useful than a more sophisticated model that cannot be clearly explained to QA, process owners or inspectors.
Not all analytical methods used in this context are, by themselves, Artificial Intelligence or Machine Learning. Correlation analysis is a conventional statistical technique, while clustering, anomaly detection and predictive risk scoring may involve Machine Learning methods depending on their implementation. In practice, these methods can be viewed as part of an analytics continuum supporting Quality Reviews.
Four approaches are especially useful in the context of Quality Reviews:
- Correlation analysis. Helps answer questions such as: “When overdue deviations increase, do operational delay hours also increase?” Correlation does not prove causation, but it highlights relationships worth investigating in PQR, APR and Management Review.
- Behavioural grouping and clustering. Observations such as months, lines or areas can first be visualised according to known operating phases or, where appropriate, grouped using unsupervised Machine Learning methods to identify patterns that were not predefined. In the illustrative case, the reactive, transition and structured-support phases are known categories used to visualise how quality and operational behaviour changes over time. In a true clustering application, the grouping method, assumptions and stability of the clusters should be defined and documented.
The illustrative dataset is intentionally small and simplified. The temporal association between structured QA-Production support and the subsequent reduction in overdue deviations and operational delays does not isolate the effect of the intervention from other factors that may have changed during the same period. The figures should therefore be interpreted as examples of signal detection and management discussion, rather than as statistical evidence of causality or effectiveness.
- Anomaly detection. Flags unusual behaviour, such as sudden jumps in deviation aging or investigation cycle time, or process areas that depart from their historical pattern. This can support earlier review before an issue becomes critical.
- Risk scoring. Estimates which open deviations are more likely to become overdue based on variables such as age, due date, missing documentation, days since last update, QA follow-ups and operational impact. The output is a prioritisation signal, not a GMP decision.
Where predictive risk scoring is used repeatedly, the threshold for action, the impact of false positives and false negatives, and the ongoing performance of the model should be defined and periodically reviewed. The model should not be considered stable simply because it performed well on the original dataset.
In all cases, the AI or ML component is used for signal detection and prioritisation, not for automated GMP decision-making. The model does not approve deviations, close CAPAs or release batches; it informs human judgement.
Other Practical Applications in Quality Reviews
The same logic can be applied beyond deviation management.
- CAPA management. AI-supported analytics can connect CAPA closure delays, effectiveness-check outcomes and recurrence of similar deviations or audit findings. This can help management distinguish isolated delays from broader signals of weak corrective-action effectiveness.
- Complaint trending. Natural language processing can group similar complaint descriptions, detect recurring terminology or emerging product-quality themes, and connect them with batch, product family, manufacturing area or deviation data. The purpose is not to classify complaints without review, but to highlight patterns that may not be visible when complaints are reviewed individually.
- OOS/OOT and analytical trends. Anomaly detection can flag unusual shifts in analytical results, investigation cycle times or repeated invalidated results across methods, products or laboratories. QC, QA and process owners can then determine whether further investigation or preventive action is warranted.
The approach can also be extended to supplier performance, audit findings and training effectiveness. In every case, the principle remains the same: AI and ML support earlier signal detection and prioritisation, while interpretation, investigation and GMP decisions remain under accountable human review.
Data Readiness Before Model Readiness
Before selecting an algorithm, organisations need reliable and consistently defined data. Due dates, status fields, process areas, investigation phases and operational-impact categories must be used consistently. Data ownership, extraction logic and transformation rules should be traceable, and the meaning of each variable should be understood by both technical teams and process owners.
If key information is stored only in unstructured comments, if statuses are not updated promptly, or if different sites use the same field in different ways, the first improvement is not a more sophisticated model. It is stronger process discipline and data governance. In many organisations, the first AI/ML project is therefore also a data-quality project.
Compliance, Data Integrity and Explainability
To keep AI usable and defensible in a GxP environment, boundaries must be clear. The first boundary is intended use: the level of control, documentation, validation and oversight should be proportionate to how the tool is used and to its potential impact on GMP decisions. This proportional approach is consistent with the risk-based principles described in ICH Q9(R1) and the Pharmaceutical Quality System framework in ICH Q101,2.
The emerging European GMP framework points in the same direction. The draft EU GMP Annex 22 on Artificial Intelligence places particular emphasis on intended use, model performance, test data, explainability and continued oversight for AI/ML models used in critical GMP applications. The concurrently revised draft Annex 11 reinforces lifecycle management, Quality Risk Management, data integrity and validated-state expectations for computerised systems. At the time of writing, both documents remain consultation drafts rather than effective Annexes, so the final regulatory texts should be checked when issued3,4.
A practical governance distinction – not a formal regulatory classification – is shown in Table 1.

Table 1. Practical intended-use categories for AI/ML-supported Quality Review activities. The categories are proposed as a governance aid and are not intended as a formal regulatory classification.
First, “a signal is not a conclusion, and a model output is not QA approval”. Exploratory analytics is not a substitute for GMP decision-making. Any action based on AI-supported insight still has to go through established deviation, CAPA, PQR/APR, Management Review and batch-release processes, as applicable. The intended use should therefore define not only what the model is expected to do, but also what it is explicitly not authorised to decide.

Figure 3. Governance boundary for AI/ML-supported Quality Review signals.
Controlled data sources feed analytics and AI/ML-supported signal detection, but the output remains subject to QA and process-owner review. Decisions remain within established GMP processes and accountable human oversight.
Second, data integrity becomes even more critical. If deviation status is not updated consistently or QA follow-ups are poorly recorded, risk scoring and pattern detection will be misleading. Controlled and traceable data extraction, clear ownership of datasets, documented transformation rules and human review of outputs are essential parts of the control strategy.
Third, explainability is more important than sophistication. A simple decision tree or clustering result that process owners can understand and challenge is more useful in a GMP context than a complex black-box model. Explainability is not only a technical preference; it is part of the quality oversight strategy.
Where a model is used on a recurring basis, organisations should also define version control, access management, change assessment, periodic performance review and retirement criteria. These controls should remain proportionate to intended use and risk.
How to Start: Small, Targeted Use Cases
The recommended way to introduce AI into Quality Reviews is not a large “AI transformation” project, but a small, controlled use case:
1. Start with a specific quality question, for example: “Why are overdue deviations increasing in Area X, and what is the operational impact?”
2. Define the intended use and state clearly which decisions remain with accountable personnel.
3. Assess data readiness, ownership, consistency and traceability before selecting a model.
4. Connect the relevant quality and operational datasets using documented transformation rules.
5. Apply an explainable analytical method and document assumptions, limitations and thresholds.
6. Review and challenge the signals with QA, process owners and relevant technical experts, including the impact of false positives and false negatives.
7. Implement actions through existing GMP processes and track whether the expected effect is sustained in subsequent reviews.
This approach helps organisations build confidence gradually: first by improving data connection and review discipline, then by introducing more advanced analytics only where there is a clear quality question and a controlled intended use.
It also keeps the focus on real problems, preserves accountability and allows the organisation to refine the method step by step.
Conclusion: More Connected, More Useful Quality Reviews
AI and Machine Learning will not replace QA judgement, and in a GxP environment they should not. Their role is to help QA teams move from isolated KPI reporting to connected, risk-based interpretation. By connecting quality and operational data, using explainable methods and keeping decisions within the GMP framework, organisations can make their Quality Reviews more than retrospective reports. They become management tools that surface meaningful signals earlier, support better questions and lead to better decisions.
From a Quality System and Compliance perspective, the main message is simple: AI should not make Quality Reviews more complicated; it should help make them more connected, more risk-based and more useful for management.
* Figures and related data are illustrative and anonymized, and should not be interpreted as Kedrion performance data or company-specific data
References and regulatory context
1. ICH Q9(R1), Quality Risk Management. Step 4, 18 January 2023.
2. ICH Q10, Pharmaceutical Quality System. Step 4, June 2008.
3. European Commission, EudraLex Volume 4, Draft Annex 22 - Artificial Intelligence. Stakeholder consultation opened 7 July 2025 and closed 7 October 2025.
4. European Commission, EudraLex Volume 4, Draft revised Annex 11 - Computerised Systems. Stakeholder consultation opened 7 July 2025 and closed 7 October 2025.
About the Authors
Simone Baisi is a Global GMP & GDP Quality Compliance Specialist and Microsoft Copilot Champion, Kedrion Biopharma, Italy.
Wolfgang Schmitt is Vice President at CONCEPT HEIDELBERG.
Go back
来源:GMP Journal 技术问答 · gmp-journal.com

