Abstract A007: Racial disparities bias oncology AI models

Abstract Deep learning-based computational pathology models can be trained to make accurate diagnostic, prognostic, and therapeutic response predictions. Such models have also recently received regulatory approval in Europe and in the US. However, a majority of these models are trained and independently evaluated on large repositories of data representative of overall patient populations without extensive evaluation on underrepresented minorities or other subgroups. Here we show that such pathology AI models can be biased when independent and external evaluation test sets are stratified by race, demonstrating the need for more extensive evaluation of such models before regulatory approvals Citation Format: Richard J. Chen, Drew F.K. Williamson, Ming Y. Lu, Tiffany Y. Chen, Jana Lipkova, Muhammad Shaban, Faisal Mahmood. Racial disparities bias oncology AI models [abstract]. In: Proceedings of the 15th AACR Conference on the Science of Cancer Health Disparities in Racial/Ethnic Minorities and the Medically Underserved; 2022 Sep 16-19; Philadelphia, PA. Philadelphia (PA): AACR; Cancer Epidemiol Biomarkers Prev 2022;31(1 Suppl):Abstract nr A007.