APRIL 20249The guiding practices are shown in the table. Each guiding practice has explanation in the related publication on how to apply it. Guiding practices 1, 2, 5, 6, 9 are straightforward, logical and good implementable advise. Guiding practice 3 makes sense, however might be in conflict with the EU-AIA. Guiding practice 4 seems logical, but might need adoption for rare diseases where only small data sets are available. Guiding practice 7, 8 and 10 deserve special attention form manufacturers and users, since they cause regular issues in practice. The MDR uses additional concepts to the GMLP such as benefits need to outweigh the risks, the medical device needs to be state of the art, and clinical evidence need to be available. However these concepts are already part of the MDR requirements.ConclusionThe EU-AIA makes it difficult or impossible for good AI-based medical devices to be placed on the EU market. The EU-AIA should be fully consistent with the MDR. This is most easily achieved, by not including the MDR in EU-AIA annex II section A, but by including the requirements of the EU-AIA in MDR MDCG guidance.Recently the USA has become the preferred location to place AI-based medical devices on the market, because of the complexities of the MDR. The additional complexity introduced by the EU-AIA will accelerate this development. There will be less choice and delayed access to crucial digital health innovation for European patients and healthcare professionals, which cannot be the purpose of the EU-AIA. The goal of Good Machine Learning Practices is to promote safe, effective, and high quality medical devices that use artificial intelligence and machine learning1. Multi-disciplinary expertise is leveraged throughout the total product life cycle.6. Model design is tailored to the available data and reflects the intended use of the device.2. Good software engineering and security practices are implemented.7. Focus is placed on the performance of the human-AI team (AI usability).3. Clinical study participants and data sets are representative of the intended patient population.8. Testing demonstrates device performance during clinically relevant conditions.4. Training data Sets are Independent of test sets.9. Users are provided clear, essential information.5. Selected reference datasets are based upon best available methods.10. Deployed models are monitored for performance and re-training risks are managed.Good Machine Learning Practices (GMLP)MDR play a major role. The proposed EU-AIA can make an unacceptable situation worse for the European patient. Currently most AI-based medical devices are placed first on the USA market and it is questionable if they will reach the EU market.When the MDR is removed from the EU-AIA annex II section A, then most inconsistencies and duplications are removed. AI-based medical devices are already in great detail regulated under the MDR. The European Commission has a scientific workgroup CORE-MD investigating what additional requirements are needed for AI-based medical devices under the MDR.Good Machine Learning PracticesThe FDA, Health Canada, and the UK MHRA have published Good Machine Learning Practices (GMLP) that can be used to develop AI medical devices. The goal of Good Machine Learning Practices is to promote safe, effective, and high-quality medical devices that use artificial intelligence and machine learning. The 10 practices identify areas where the International Medical Device Regulators Forum (IMDRF) and international standards organizations could work to advance GMLP. The guiding practices can also be used to develop MDCG guidance for the EU-MDR, instead of using the EU-AIA.
< Page 8 | Page 10 >