A clinical model reflects the populations, equipment and practices in the data it was trained on, and can perform measurably worse for groups those data under-represent. This is a safety problem rather than a fairness footnote: a model that misses disease more often in one group produces worse care for that group, quietly and at scale.
We report bias alongside performance because the two are not separable, and because degradation over time — as populations and practice change — is the same problem observed later.
Artificial intelligence in healthcare: where AI is used, what the evidence shows, regulation of AI medical devices, bias and safety risks, and governance requirements.
How AI supports medical diagnosis and clinical decision-making: imaging, pathology, ECG, risk prediction, validation requirements, and clinician oversight.