There are many strategies that can be implemented in the development of AI technology to make predictions more equitable. These bias mitigation strategies can used during data pre-processing before training a model. The strategies can also affect the training process of an algorithm, such as minimizing error rates across protected classes. In this i.equalcare initiative, we will test the potential of different techniques in AI models used for clinical purposes such as disease diagnosis and risk prediction.
Research Projects
Addressing Bias with Data Augmentation
In this initiative, we look at ways to address imbalanced data. This is a prevalent issue in medical datasets as there are often minority groups in populations (based on race/ethnicity, gender, sexuality, etc.) that are underrepresented. Additionally, some regions may not be as prevalent in data because they lack the extensive electronic health systems and resources required to collect samples such as medical images. Using data augmentation techniques, we can achieve fair performance of AI models for all demographics.
i.equalcare 