Post by Colin Mitchell

💥 Key considerations for use of synthetic data in the development of AIaMD💥 A privilege to have worked with Puja and the rest of the MHRA team/Expert Working Group on a series of discussions and drafting of a guide for manufacturers and approving bodies assessing their use of synthetic data (SD) in the development of AI as a Medical Device (AIaMD). It is not a long document (and contains a glossary) so please take a look but here are some headlines: ☂️An overarching question is: Can you justify your approach and demonstrate that you have minimised risks as far as possible, appropriate to the AIaMD use case? ⚠️ Synthetic data should be part of a broader evidence strategy and justified based on specific context (potential justifications include boosting datasets to addres bias or where real data are unavailable) 🔎 The regulatory acceptability of synthetic data rests on three overarching principles: fidelity, representativeness, and transparency and there are a range of specific key considerations that fall within these for synthetic data (see report for details!) ❯❯❯❯ Post-market surveillance may be particularly important where synthetic data have been relied on- to address uncertainty and sustain regulatory confidence In addition to the Expert Working Group, many thanks to Valena Reich for her brilliant work on this as part of the PHG Foundation team! Dominique Chu, Alastair Denniston, Prof Alejandro Frangi FREng, Ibrahim Habli, Marinos Ioannides, Puja Myles, Johan Ordish, Mark Palmer, M.D., Ph.D., Russell Pearson, Gavin Quigley, Allan Tucker, Christopher Yau, Elizabeth Redrup Hill, PhD