The International Medical Device Regulators Forum (IMDRF) has published Machine Learning-enabled Medical Devices: Key Terms and Definitions (IMDRF/AIMD WG/N67, Edition 1). This foundational guidance establishes a common vocabulary for artificial intelligence (AI) and machine learning (ML) in the medical device sector. Its purpose is to create uniform expectations and understanding, improve patient safety, inspire innovation, and encourage access to breakthroughs in healthcare technology.
Artificial intelligence is broadly defined as the use of algorithms or models to perform tasks, make decisions, or generate predictions. Within AI, machine learning is a subset where models are trained on data, enabling them to learn patterns without explicit rule-based programming. The IMDRF document situates these concepts within a regulatory and clinical context, ensuring clarity when applied to medical devices.
One of the key goals of the guidance is to reduce confusion across jurisdictions. Manufacturers, regulators, and clinicians may use different terms for the same concepts, such as “model,” “training,” or “retraining.” This lack of alignment can complicate regulatory submissions and reviews. The IMDRF’s definitions create a standard set of terms that can be consistently referenced across regulatory frameworks and development programs.
In February 2025, IMDRF released Good Machine Learning Practice (GMLP), which builds on the definitions in N67 by providing ten guiding principles for the development, validation, and monitoring of ML-enabled devices. The link between the two documents is crucial: N67 defines the language, while GMLP sets expectations for practice across the product lifecycle.
Key Terms and Their Impact
The N67 guidance defines terms such as “model,” “training set,” “test set,” “drift,” “bias,” “retraining,” “locked model,” and “continuous learning.” These definitions are not academic—they directly influence how safety risks are assessed, how validation studies are structured, and how regulatory change control is applied. For instance, a locked model is one that does not evolve after deployment, while a continuous learning model adapts over time, requiring additional oversight and safeguards.
Understanding drift and bias is particularly important. Drift refers to performance degradation when the underlying data distribution changes, while bias indicates systematic error or unequal performance across patient subgroups. The IMDRF document clarifies these terms to support manufacturers in identifying when retraining or remediation is required to maintain safety and performance.
Why Uniform Definitions Matter
A harmonized vocabulary enhances regulatory predictability and cross-border alignment. With common definitions, manufacturers can prepare more consistent submissions, and regulators can apply more transparent and standardized review processes. It also helps notified bodies, standards committees, and audit organizations maintain consistent evaluation criteria.
Clear definitions are equally important for clinicians and patients. When a device is described as “continuously learning,” stakeholders need to understand the precise boundaries of its adaptation. This clarity reduces risks of misinterpretation that could compromise patient safety or compliance.
Integration with Regulatory Practice
The IMDRF’s N67 definitions are now referenced in the GMLP principles adopted by multiple regulators, including those in the U.S., UK, EU, and Canada. This reinforces the importance of shared terminology as the basis for regulatory policy. Together, N67 and GMLP create a roadmap for the development and oversight of AI/ML-enabled devices, from design and testing to monitoring and lifecycle management.
Implications for Developers
Manufacturers must integrate IMDRF definitions into their development practices from the outset. Risk management plans, validation strategies, and change-control procedures should explicitly reflect terms such as drift, retraining, and continuous learning. Clinical performance evaluation must be designed using clearly defined training and test sets, while monitoring strategies must track performance shifts aligned with N67 definitions.
Failure to align with this common vocabulary can lead to misinterpretation, regulatory delays, or gaps in safety oversight. By embedding these terms into development and documentation, companies can demonstrate compliance and strengthen the credibility of their devices.
Conclusion
The IMDRF’s Machine Learning-enabled Medical Devices: Key Terms and Definitions guidance represents a milestone in harmonizing global understanding of AI/ML in healthcare. By defining key terms such as model, drift, and retraining, it lays the foundation for safe innovation and regulatory clarity. Together with the GMLP framework, it provides a roadmap for developers and regulators alike as AI-enabled healthcare technologies continue to evolve.
If your team is developing an ML-enabled device and needs support in aligning with IMDRF definitions and regulatory expectations, contact MDx CRO to discuss how we can guide your strategy from concept to approval.
