Free Podcast: Q&A session on requirements for SaMD and AI/ML in medical devices

Val Theisz answers questions posed by JMDR subscribers on the regulatory requirements for SaMD and AI/ML in medical devices. The slides she uses are also available for free download alongside the audio and video files.

Tune in to our FREE podcast!

In this interview, hear Val answer the following questions:

1. Are there particular challenges faced by industry by having potentially different regulatory requirements applied by various regulatory authorities around the world? What are those challenges and how are they being addressed?
2. What is your opinion on the EU approach to AI and the recently released draft regulation? How does the EU’s approach differ from the rest of the world (particularly China, the USA and UK)?
3. How do you see regulators keeping up with the fast-moving pace of the development and implementation of these types of technologies? Are the technologies changing faster than the regulatory framework? If so, please explain, and how are industry and regulators dealing with that reality?
4. These types of devices have the potential to transform healthcare and improve patient care significantly. What are some of the most significant potential benefits of these types of technologies? What are some of the potential pitfalls, risks and “watch outs” in terms of how these technologies are being used or may be used in the future?
5. What do you consider as best practices for monitoring post-release evolving AI? What would be your suggested methodologies for ensuring safety and performance while AI is evolving?
6. What are your views on continuous learning systems and how these systems should be addressed by regulatory frameworks? How can a manufacturer obviate the need to keep updating the technical documentation with each iteration and software upgrade?
7. For AI/ML, can you give any insights (and examples) on what Notified Bodies are accepting as sufficient for Post-Market Surveillance or Post-Market Clinical Follow-up?
8. To what extent can, and should, a manufacturer align the approval process of medicinal products and software as a medical device when they are meant to be used together? Is the current system in Europe which provides for separate approval processes by separate regulatory entities – the European Medicines Agency or a national Competent Authority versus Notified Bodies – a hurdle?
9. What standards should be used for AI/ML? Also, are there any standards that should be followed for cybersecurity now?
10. What are the procedures as well as the consequences of a permanent discontinuation or interruption in the manufacturing of a device of this type that is likely to lead to a meaningful disruption in the supply of that device considering the new European regulatory system?
11. For software as a medical device with AI/ML, what should be the appropriate performance targets? For example, should the targets be set according to (a) guidelines, (b) the “gold standard” of a clinician’s performance on the same task, (c) a competitor’s performance targets for the same or similar tasks or (d) something else?
12. For AI/ML, a clear description of methodology for development should be available. Do you consider this a procedure in the QMS, or are requirements, a validation plan and documentation sufficient?
13. Again for AI/ML, are there guidelines on the size of data sets? Or only on statistics?
14. How should the risks of over/under reliance on AI be addressed by manufacturers/providers?
15. The final question is very specific and relates to the US regulations. The question is: We have a locked algorithm (same input, same output) using automatic iterative mathematical processes, which supports diagnosis (on ultrasound units), and which has been developed and tested on a sufficient and specific (ideally non-biased) patient database. This algorithm can be executed in all typical clinical applications of echographers (from cardiovascular to obstetrics/gynaecology) but does not directly provide a final diagnosis and only performs calculations useful for the image interpretation. Would it be sufficient to include complete clinical evaluation/validation in a 510(k) submission to obtain FDA clearance? Furthermore, to what extent is a description of the “nature” of the algorithm required? Finally, is there a need for periodic review for both open and locked algorithms?