Disparities in memorisation bias across patient subgroups

Determine whether disparities in membership-inference risk across patient groups, particularly the disproportionate privacy burden affecting groups underrepresented in training data, also extend to memorisation bias in medical AI models.

Background

The study does not perform a subgroup analysis of memorisation bias. It cites prior work showing that membership-inference risk is unevenly distributed, with patients from groups underrepresented in training data bearing a disproportionate share of the privacy risk.

The unresolved question is whether these previously observed disparities in membership-inference risk also occur for longitudinal memorisation bias, which can alter predictions on the future records of patients whose historical data were used for model training. Resolving this issue would clarify whether the diagnostic and privacy consequences identified in the paper are concentrated among particular patient populations.