Bias in Clinical AI
Also known as: Algorithmic Bias
Systematic differences in an AI model's accuracy across patient subgroups, often traced to unrepresentative training data.
Bias in clinical AI refers to systematic differences in an AI model's accuracy or output across patient subgroups, such as by age, sex, race, or the equipment used to capture the data, most often traced back to training data that underrepresented those subgroups.
Bias can be present even when a model's overall (aggregate) accuracy looks strong, since aggregate performance can mask weaker performance in a smaller subgroup. Subgroup performance reporting is one reason validation studies increasingly break results out by population.