The least glamorous bottleneck in European health AI is not the consent banner, the model card, or the policy panel. It is the dull question no demo answers cleanly: what evidence would make this claim believable to a regulator, a hospital, and a patient at the same time? A new Comment in Nature Biomedical Engineering puts that problem in institutional language. Europe, it argues, may need regulatory science infrastructure as much as it needs another layer of guidance. ## The centre proposal is not a deadline Nature Biomedical Engineering describes the proposal from Anett Schönfelder, Alastair K. Denniston, Alejandro F. Frangi and others as a call for Centres of Excellence in Regulatory Science and Innovation for medical devices and medical artificial intelligence in Europe. The Comment says those centres would provide research on concepts, early evidence generation, and horizon scanning for emerging needs and technologies, according to the Nature preview. Translation for builders: this is not a new filing obligation, but it is a warning that late-stage compliance theatre is a poor substitute for evidence design. The European Medicines Agency has already been moving in this direction in medicines. In its 2021 Annual Report, EMA said its Regulatory Science Strategy to 2025 was published in March 2020 and translated into actions and work planning spanning until 2025. The same report said actions included the ACT EU initiative, focus on novel manufacturing techniques, advanced therapies and personalised medicines, investment in real-world evidence capabilities, digital innovation initiatives, and health threat work linked to the pandemic response. That timeline matters because regulatory science is not a slogan for being nicer to innovators. It is the boring machinery that tests whether regulatory categories, evaluation methods, evidence standards, and surveillance tools can cope with new technology. If Europe builds centres around that work, the practical effect is not fewer rules. It is earlier answers to the questions that otherwise arrive during approval, procurement, or enforcement. ## Why guidance cannot carry the whole load The European Commission already has an innovation lens for lawmaking. Its Research and Innovation department says the Innovation Principle requires the innovation dimension to be considered when preparing and implementing EU legislation, while upholding Europe’s standards, and says Tool 22 guides Commission services in analysing the interaction between EU legislation and innovation during impact assessment. That is useful, but impact assessment is upstream of a builder trying to validate a model, document clinical performance, or decide whether a dataset is adequate. CEPS made the older version of the same point in its Special Report No. 96 by Jacques Pelkmans and Andrea Renda. The report said the question of whether EU regulation hinders or stimulates innovation is common, but systematic analytical literature is limited and fragmented evidence or anecdotes dominate the debate among EU decision makers and business. Put less politely: many policy fights still run on examples, not measurement. Centres of excellence would not abolish disagreement, but they could give regulators and companies a shared evidence bench. ## What builders should do before the lawyer arrives EMA’s Regulatory Science Research Needs 2025 update, dated 14 July 2025, lists research needs to improve clinical research and regulatory system evolution, and frames them as opportunities for researchers. For AI and regulated software teams, that points to a concrete operating change. Build an evidence plan when the product claim is written, not when the regulatory submission is being assembled. The Nature Comment’s emphasis on early evidence generation is the useful part for product teams. If your system claims to support diagnosis, triage, monitoring, or clinical workflow, the compliance question is not only which regulation applies. It is whether your evaluation method matches the claim, whether the data supports the intended population and setting, and whether you can explain performance drift after deployment. Your vendor contract can say many comforting things; it cannot manufacture missing evidence after the pilot fails. ## Compliance tools are useful, but they are not infrastructure BusinessEurope’s repository on the impact of EU regulation on innovation lists robust scientific evidence and stakeholder engagement among criteria for regulation that can stimulate innovation. EFPIA, writing about pharmaceutical regulation, similarly argues for a predictable, clear, and consistent environment while continuing to rely on the precautionary principle. These are not fringe requests. Regulated industries are asking for rules they can build against, but also for methods that make safety and performance review less improvisational. That is where the centre idea is more interesting than another compliance guide. RegTech can help track obligations, produce reports, and reduce administrative drag. It does not decide what counts as adequate validation for a new medical AI use case, or how to compare real-world evidence across institutions. Regulatory science centres would sit closer to that harder layer: shared methods, horizon scanning, and evidence standards that can be tested before companies burn capital on the wrong proof. For readers building in AI, health tech, or regulated software, the takeaway is simple and inconvenient. Treat regulatory science as product infrastructure. Watch whether the centre proposal gets institutional backing, funding, and routes for smaller builders to access methods they could not develop alone. The winners will not be the teams with the thickest compliance memo. They will be the teams that can show their evidence before someone asks for it. ## Sources - European Union innovation needs regulatory science excellence centres

Sources