EMA AI Reflection Paper: Introduction & General Principles
How the European Medicines Agency frames AI and machine learning in medicines regulation, and the risk-based principles that apply across the lifecycle.
Data are generated and used increasingly across sectors, including the lifecycle of medicines. In healthcare, data are captured in electronic format on a routine basis. Artificial intelligence (AI) — systems displaying intelligent behaviour by analysing data and taking actions with some degree of autonomy to achieve specific goals — is an important part of the digital transformation enabling increased use of data for analysis and decision-making.
Such systems are often developed through machine learning (ML), where models are trained from data without explicit programming. However, as these technologies often use exceptionally large numbers of trainable parameters arranged in non-transparent model architectures, new risks are introduced that must be mitigated to ensure patient safety and the integrity of clinical study results. As the approach is inherently data-driven, active measures must be taken to avoid integrating bias and to promote AI trustworthiness.
Purpose of the reflection paper
This reflection paper provides considerations on the use of AI and ML in the lifecycle of medicinal products, including development, authorisation, and post-authorisation. Given the rapid development in this field, the aim is to reflect on scientific principles relevant for regulatory evaluation when these emerging technologies are applied to support safe and effective development and use of medicines.
It is crucial to identify aspects of AI/ML that fall within the remit of EMA or National Competent Authorities, as the level of scrutiny during assessment depends on this remit. The paper focuses only on AI in the medicinal product lifecycle. Medical devices with AI/ML technology may be used in clinical trials to generate evidence for a marketing authorisation application, or combined with a medicinal product — in such cases EMA assesses whether the device characteristics are adequate to generate supporting evidence.
General considerations
AI and ML tools can, if used correctly, effectively support the acquisition, transformation, analysis, and interpretation of data within the medicinal product lifecycle. Many recommendations, best practices, and learnings from model-informed drug development and biostatistics also apply to AI/ML.
Risk-based approach
A risk-based approach for development, deployment, and performance monitoring allows developers to proactively define risks to be managed throughout the AI/ML tool lifecycle. Risk includes, but is not limited to, regulatory impact. System malfunction or degradation of model performance can range from minimal to critical or even life-threatening, depending on the AI technology, context of use, and degree of influence the AI exerts.
Marketing authorisation applicants or holders planning to deploy AI/ML are expected to consider and systematically manage relevant risks from early development to decommissioning.
Early regulatory interaction
If an AI/ML system is used in medicinal product development, evaluation, or monitoring, and is expected to impact — even potentially — the benefit-risk of a medicinal product, early regulatory interaction such as qualification of innovative development methods or scientific advice is advised. The level of scrutiny depends on the risk and regulatory impact posed by the system.
Applicant responsibility
It is the responsibility of the marketing authorisation applicant or holder to ensure that all algorithms, models, datasets, and data processing pipelines used are fit for purpose and in line with ethical, technical, scientific, and regulatory standards as described in GxP standards and current EMA scientific guidelines. These requirements may in some respects be stricter than standard data science practice.
For all requests for advice or opinions, the applicant or holder is expected to provide a scientific base along with sufficient technical details to allow comprehensive assessment of any AI/ML systems used, the integrity of data, and generalizability of models to the target population and specific context of use.
Ready to test your knowledge?
Take a short quiz based on this article to check your understanding.
Take the quiz