Bias, privacy and biomedical AI
A dataset can underrepresent groups and produce uneven errors.
Loading video…
A dataset can underrepresent groups and produce uneven errors. Audit performance by meaningful subgroups and report uncertainty rather than assuming aggregate accuracy is fair. Protect identifiers, minimize data access and use consented or synthetic educational data. Leakage from patient identity or future measurements can make a model appear successful without clinical value.
Worked example
A classifier trained at one hospital performs well overall but poorly for a minority subgroup. Report subgroup counts and uncertainty, and seek external validation before use.