EC Issues Annexure 22 – Draft Guidance For AI Systems Used in Manufacture of Medicinal Products

The European Commission (EC) has published a draft guidance – Annexure 22 – Artificial Intelligence – for GxP computerised systems which interface with Artificial Intelligence models to predict or classify data with direct impact on patient safety, quality and data integrity. For example (but not limited to) tasks or decision outcomes which are being automated like “Accept / Reject” outcomes, identifying specific characteristics in material or product like type and severity of defects.

The simple to understand and practical guidance is applicable for medicinal products and active substances and applies to machine learning (AI/ML) models which obtain their functionality through training with data rather than explicitly programmed. The guidance is applicable to static models with a deterministic output which when given identical inputs provide identical outputs.

Exclusions

This guidance is not applicable to dynamic models which may adapt their performance during use by incorporating new data, models which continuously learn and adapt, probabilistic models and to Generative AI and Large Language Models (LLM) and explicitly state that such models should not be used in critical GMP applications.

The document emphasis close cooperation between all stake holders  – subject matter experts (SMEs), quality assurance, data scientists, IT and consultants during all phases of the implementation of such systems – algorithm selection, model training, validation, testing and operation which should be sufficiently documented. The model’s intended use and specific tasks it is designed for must be clearly described, along with a comprehensive characterization of the input sample space.

Testing and Test Data and Independency

The guidance sets out detailed requirements for testing of the model and acceptance criteria and lays particular emphasis on the Test Data Independency.  

Test data should cover the full sample space of the intended use, reflect complexity and all common and rare variations, be sufficient in size to calculate the test metrics with adequate statistical confidence. Any preprocessing or cleaning of the test data should be documented and justified. Sufficient test metrics should be defined including confusion matrix to measure performance of the model, sensitivity, accuracy and precision.  Acceptance criteria should be defined by a process subject matter expert (SME), documented, reviewed and approved before acceptance testing. Acceptance criteria should be at least as high as the process the AI system / model is replacing. Independency of test data should be ensured and data which is used to test a model should not have been used during development, training or validation of the model. If the test data is split from a complete pool of data from which data was used for training and development of the model, it should be ensured that the employees involved in development and training of the model never had access to the test data.

Testing must follow a documented test plan and test script, with all documentation retained and access-controlled, including audit trails. The guidance also specifies feature attribution during testing, that is the system should capture and record features of the test data that contributes to particular classification of decision (acceptance/ rejection). These should be reviewed while approving test results. The guidance also recommends logging the confidence score for each prediction or classification outcome with an appropriate threshold setting so that the model flag the outcome as ‘undecided’ rather than make unreliable predictions.

After testing, deployment of the system must proceed via change control and Quality Risk Management (QRM). The model should be placed under configuration control before deployment in operation to control and prevent unauthorised changes. There should be a system of regular performance monitoring of the model as well as input space monitoring to verify the input data are still within the model sample space and intended use.

The guidance offers a robust matrix to define, design, test, implement and for continuous oversight of AI supported computerised systems used in the manufacturing of medicinal products.

EC Annexure 22: Artificial Intelligence (July 2025)