Artificial intelligence-enabled tools for diagnostic imaging approved under old and new EU regulatory frameworks - differing standards of evidence but continuing availability
Y Ren, D Windecker, A Fraser, G C M Siontis, E G CaianiAbstract
Introduction
The EU regulatory framework for medical devices underwent a substantial transition when the Medical Device Directive (MDD) was replaced by the Regulation (MDR), starting 26/05/2021. The MDR has more stringent requirements for approval and post-market surveillance especially for software, including Medical Device Artificial Intelligence (MDAI), with new criteria for risk classification (Annex VIII, Rule 11).The largest proportion of MDAI tools is used in diagnostic imaging (DI) as a standard part of daily clinical workflows in many departments. However, MDAI classified under the MDD risk rules (legacy devices) will continue to coexist until 31/12/2028 on the EU market with MDAI approved under MDR, thus posing confusion on the final users about the underlying level of scrutiny.
Purpose
The study aims at describing the risk class distribution and associated level of evidence for currently available MDAI solutions in DI in the EU market, across regulatory frameworks (MDD vs MDR).
Methods
Information on MDAI in DI approved in the EU up to 30/11/2025 was retrieved from a publicly available database (Health AI Register) by web scraping. The association between EU regulatory pathway, MDAI risk class, and existing clinical evidence (at least 1 peer reviewed publication) was assessed by two-sided Fisher’s exact test.
Results
Among the retrieved 214 MDAI solutions, 78 (36%) were approved under the MDD and 136 (64%) under the MDR. A significant difference (p<0.01) was observed in the risk class distribution, with MDD-approved MDAI mainly in Class IIa (medium risk, n=42; 54%) followed by Class I (low risk, n=36; 46%), and no Class IIb (medium-high risk). Conversely, MDR-approved MDAI were predominantly in Class IIa (n=92; 67.6%) or Class IIb (n=43; 31.6%), with only n=1 (0.8%) class I device. As a result, 36/214 (17%) MDAI solutions currently available on the EU market are legacy self-declared Class I software products, without notified body involvement. An overall association between device risk class and evidence type was observed, significant under the MDR (p<0.01), with Class IIb devices more often supported by peer-reviewed publications (34/43; 79%) than Class IIa devices (42/92; 46%).
Conclusions
With MDR,a shift in MDAI risk classification distribution towards higher classes is visible, implying higher scrutiny and more clinical evidence, as noted for Class IIb. However, Class I legacy MDAI DI solutions are still available, despite they could face higher, risk-based classification and scrutiny under the MDR. Additional manufacturers’ obligations for high-risk AI (i.e., all MDAI as defined by the EU AI Act, increasing their performance and safety), will enter in force only on 2/8/2027. Healthcare professionals using MDAI in DI need access to information about the regulatory framework approval, risk class and existing evidence when selecting tools for purchase or interpreting clinical results.Figure.