Hospitals can now measure how AI brain tumor models fail specific patients

Researchers built a new open-source tool to check if AI medical models work equally well for all patients. It turns out that patient characteristics, not the AI model itself, explain most of the performance differences.

US health regulators have approved over 1,000 AI medical devices. But few of these devices have formal checks for fairness across different patient groups. This new tool means hospitals can now identify specific biases in AI models before they affect patient care.