Training data is not the only factor that can make AI tools less accurate for some groups of patients. The choice of model can also play a role. This study compared four AI models used to outline the structures of the heart on cardiac MRI scans, using data from the UK Biobank and varying the proportion of men and women, and of Black and White people, in the training data.
Three of the four models showed bias by sex, with better results for men when men were the majority or when the training data was balanced. All four models showed bias by race. In general, increasing the share of a group in the training data improved the model’s performance for that group, but the size of this effect varied considerably from one model to another.
The authors conclude that representation of different groups in training data matters, but that the choice of model also strongly affects bias, so each model needs to be carefully evaluated before use.
Learn more about this study here: https://doi.org/10.1007/978-3-031-45249-9_21
Reference:
Lee, T., Puyol-Antón, E., Ruijsink, B., Aitcheson, K., Shi, M., & King, A.P. (2023). An Investigation into the Impact of Deep Learning Model Choice on Sex and Race Bias in Cardiac MR Segmentation. In S. Wesarg et al. (Eds.), Clinical Image-Based Procedures, Fairness of AI in Medical Imaging, and Ethical and Philosophical Issues in Medical Imaging (Lecture Notes in Computer Science, Vol. 14242). Springer.