Principally, this means we need to ensure:
- There are no direct join keys (e.g., timestamps, time-deltas, code names that correspond to a small number of patients, etc.). Code names are currently discarded so that point is not an issue remotely here, but it is worth noting in general.
- Continuous / non-categorical distributions are used where possible to ensure support is not concentrated on only real data.
- High-precision outputs are noised.
Note that none of these are truly serious concerns here, as there is no real data relating to medical codes or laboratory results preserved, and all aspects of the synthetic patients' record are generated independently (so it is literally impossible even with the current approach to infer anything about an existent real patient given a synthetic patient), but to avoid even the possibility of the appearance of a membership inference attack, it is best to be extra careful.
Principally, this means we need to ensure:
Note that none of these are truly serious concerns here, as there is no real data relating to medical codes or laboratory results preserved, and all aspects of the synthetic patients' record are generated independently (so it is literally impossible even with the current approach to infer anything about an existent real patient given a synthetic patient), but to avoid even the possibility of the appearance of a membership inference attack, it is best to be extra careful.