Study reveals patient characteristics matter more than AI model design for brain tumor segmentation accuracy

Researchers evaluated 18 open-source brain tumor segmentation models across 648 glioma patients to assess fairness and equity in AI medical devices

Patient identity factors consistently explained more performance variance than the choice of AI model itself

Clinical variables like molecular diagnosis, tumor grade, and surgical extent predicted segmentation accuracy more strongly than model architecture

Voxel-wise spatial analysis identified neuroanatomically localized biases that were specific to different brain regions but often consistent across multiple models

Study highlights the need for formal equity assessments in AI medical devices, despite over 1,000 FDA-authorized AI medical devices currently in use