Introducing artificial intelligence (AI)-based medical devices to the European market demands more than cutting-edge technology. Achieving compliance with the EU Medical Device Regulation (MDR 2017/745) hinges on a comprehensive understanding of the regulatory landscape governing product safety, clinical performance, and market approval. Notably, the clinical data requirements necessary for CE marking represent some of the most critical and challenging aspects of this process.
For manufacturers and regulatory professionals, it is imperative to provide clear and compelling evidence that AI-based medical technologies are not only safe and effective but also deliver clinically meaningful results in authentic healthcare settings. This process begins with establishing robust criteria for what constitutes valid clinical data under the MDR framework. It further requires careful consideration of when prospective clinical investigations are necessary to validate device performance, as opposed to when alternative forms of evidence, such as retrospective data analyses, real-world clinical data, or published literature, may be deemed acceptable by regulators. Successfully meeting these expectations demands a strategic approach to evidence generation, tailored to the intended use and risk classification of each device.
A structured and evidence-driven approach to clinical data requirements ensures that each AI medical device is supported by transparent, verifiable, and scientifically credible documentation. This framework is essential for obtaining CE marking, maintaining regulatory confidence, and ensuring the long-term integrity of AI medical technologies in European healthcare.
Understanding clinical data requirements
Under the MDR, manufacturers must demonstrate that a device meets both safety and performance requirements. Clinical data refers to all documented information that shows how a device impacts human health, whether through its safety profile, performance characteristics, or overall benefit to patients. The MDR defines clinical benefit as “the positive impact of a device on the health of an individual, expressed in a meaningful, measurable, patient-relevant clinical outcome, including outcomes related to diagnosis or to a positive impact on patient management or public health.”
In practice, this means manufacturers must go beyond demonstrating technical performance.
They must provide evidence that their device contributes to better health outcomes or patient management in measurable ways. The quality, relevance, and traceability of that evidence determine the robustness of the Clinical Evaluation Report (CER) submitted to the Notified Body.
Why clinical data for AI devices is complex
AI medical devices present unique challenges in meeting clinical data requirements and generating credible clinical evidence. Many AI-based systems, such as those used for image interpretation, diagnostic support, or predictive analytics, do not directly interact with patients. Instead, they analyze data, assist clinical decisions, or provide recommendations to healthcare professionals. These devices produce what regulators call an indirect clinical benefit rather than a direct therapeutic one.
For example, an AI application that measures cardiac structures on MRI or echocardiography does not treat a patient directly. Its value lies in improving diagnostic accuracy and speed, which enhances patient care indirectly. Demonstrating this indirect benefit in line with clinical data requirements under the EU MDR requires a clear link between improved diagnostic performance and meaningful clinical outcomes.
The MDR allows such devices to rely on indirect clinical evidence, but the burden of proof remains high. Manufacturers must show that the AI system’s outputs are reliable, relevant, and compliant with all applicable clinical data requirements, leading to improved clinical decision-making.
When clinical investigations are required
High-risk medical devices, such as implantables or systems that directly affect the human body, typically require new clinical investigations involving human subjects. The MDR defines a clinical investigation as “a systematic investigation involving one or more human subjects to assess the safety or performance of a device.”
For AI devices, the decision to conduct a clinical investigation depends on three factors: the level of risk, the novelty of the technology, and the connection between the device’s outputs and clinical outcomes.
If the AI model directly influences treatment, diagnosis, or prognosis, and its effects on patient care are not yet validated, then a clinical investigation is required to confirm safety and performance.
Conversely, if the device performs well-defined analytical functions without direct patient impact, and its algorithms are validated through strong datasets, regulators may accept alternative forms of evidence. The key is to justify this decision within the clinical evaluation plan and ensure that every assumption is supported by data.
When alternative evidence is acceptable
Many AI medical devices can rely on alternative evidence sources if their use case involves indirect benefit and minimal patient risk. The MDR permits the use of retrospective validation, literature data, and non-clinical studies, provided they are scientifically justified and transparently documented.
Retrospective validation involves using existing clinical data to demonstrate accuracy and reliability. For instance, an AI model that quantifies cardiac function from archived imaging data can be validated retrospectively by comparing its results with expert assessments.
Published literature can support indirect clinical benefit when existing studies show that improved diagnostic precision or reduced analysis time contributes to better patient management. If literature establishes a proven connection between enhanced workflow and improved health outcomes, this evidence can be integrated into the clinical evaluation.
Non-clinical evidence may be acceptable under Article 61(10) of the MDR when human studies are not appropriate. For devices that perform limited or purely analytical functions, bench testing, algorithm performance validation, and usability studies can demonstrate safety and performance.
The manufacturer must provide a detailed scientific rationale explaining why human clinical data are unnecessary and how the non-clinical data sufficiently demonstrate clinical benefit.
Building a clinical evidence portfolio
To comply with EU MDR clinical data requirements, manufacturers should develop a complete and well-documented clinical evidence portfolio that includes the following elements.
1. Intended purpose and claims – clearly define the medical purpose of the device and the claims it makes. The intended use must be precise, measurable, and aligned with the device’s actual functionality.
2. Comprehensive source list– document all sources of clinical data, including investigations, retrospective studies, peer-reviewed literature, post-market data, and usability assessments. Evaluate each source for reliability and relevance.
3. Systematic literature review – conduct a systematic review according to MEDDEV 2.7/1 Rev. 4 to ensure the inclusion of high-quality evidence and to compare the device with equivalent technologies.
4. Performance and safety summary – provide a clear synthesis of all available data, summarizing clinical outcomes, usability testing, and any observed adverse events.
5. Benefit-risk analysis – demonstrate that the benefits of the device outweigh its potential risks, considering both clinical and technical factors.
6. Post-market surveillance and follow-up – establish a continuous post-market surveillance (PMS) and post-market clinical follow-up (PMCF) plan to monitor real-world performance and safety.
Us2.ai cardiac measurement software
A strong real-world illustration of these principles is Us2.ai, a CE-marked AI software that automates the analysis of cardiac structures on echocardiography and MRI. The platform uses deep learning to measure ventricular volumes, ejection fraction, and wall motion with precision comparable to expert cardiologists.
Us2.ai achieved CE marking under the MDR through a combination of retrospective validation studies, published literature, and extensive algorithm performance benchmarking. Instead of conducting new prospective clinical trials, the manufacturer demonstrated compliance with the EU MDR clinical data requirements by validating the AI model on thousands of archived echocardiographic datasets from diverse clinical sites and patient populations.
The evidence showed that automated measurements produced by the software were statistically equivalent to expert manual assessments and that their use significantly reduced analysis time, enabling faster and more consistent diagnostic workflows.
Peer-reviewed studies further confirmed that this automation improved reproducibility, reduced operator dependency, and supported better clinical decision-making in real-world cardiology practice.
This approach illustrates how AI medical devices delivering indirect clinical benefit can still fully satisfy clinical data requirements through scientifically justified, transparent, and well-documented evidence strategies. By combining retrospective clinical data with published outcomes and usability insights, the manufacturer successfully demonstrated both safety and clinical performance without the need for a new human clinical investigation.
Best practices for AI manufacturers
AI device manufacturers can meet MDR clinical data requirements more effectively by following a structured, evidence-driven approach. Begin with a clearly defined intended purpose and measurable claims. Collaborate with clinical and regulatory experts to ensure that the data strategy aligns with MDR expectations.
Ensure that retrospective datasets accurately represent the intended user population and that all validation activities are traceable and reproducible. Include usability and real-world testing data, not only laboratory results, to show performance under clinical conditions.
Maintain up-to-date documentation throughout the product lifecycle.
Implement a proactive post-market surveillance plan that gathers real-world evidence and quickly identifies safety or performance signals. For adaptive algorithms, establish a controlled process for software updates, ensuring that every version maintains validated performance and safety profiles.
Conclusion
The future of MDR compliance for AI will depend on the ability to transform algorithmic performance metrics into clinically meaningful endpoints. Manufacturers should integrate continuous learning systems, human oversight, and post-market analytics into their evidence plans, ensuring that updates to models remain clinically validated and explainable. This approach not only meets clinical data requirements but also anticipates the growing regulatory focus on transparency and lifecycle monitoring for adaptive AI systems.
Clinical data should therefore be treated as a dynamic asset, one that builds trust among Notified Bodies, accelerates CE marking, and enhances partnerships with healthcare institutions.
By embedding evidence generation into product design and lifecycle management, manufacturers can create a defensible and scalable regulatory position.