Algorithmic Bias (Healthcare)
Also known as: AI bias in healthcare, model bias, health equity AI
Systematic difference in an AI model's performance across patient groups, usually introduced through unrepresentative training data or a target variable that stands in poorly for the thing being predicted.
Algorithmic bias in healthcare rarely originates in intent. It typically enters one of two ways. First, through training data in which some patient groups are thin, so the model has learned less about them and performs worse on them. Second, and harder to detect, through proxy targets: the model predicts something measurable that stands in for the clinical concept of interest, and the proxy carries historical inequity the clinical concept does not. The widely cited example is the use of prior healthcare spending as a proxy for health need, which systematically under-identified Black patients because less had historically been spent on their care. Detecting this requires asking what the target variable actually measures, not only reviewing aggregate accuracy.