Machine learning (ML) has profound potential to make healthcare more accurate, accessible, and personalized. Today, ML models developed with medical images are being validated in clinical settings to expedite diagnosis in ophthalmology and radiology. Electronic health records (EHRs) are also being utilized to build ML systems for more inclusive clinical trials design so that study findings translate to clinical practice.
Progress withstanding, ML in healthcare faces significant ethical and methodological barriers. Ethically, it is evident that ML systems can reinforce inequity by perpetuating biases intrinsic to clinical data or poor design choices. Substantial differences in the accuracy of ML models across protected attributes such as race have been observed in numerous applications, recently in a state-of-the-art radiograph diagnostic model. To prevent ML from exacerbating inequity in healthcare, there is an urgent need to precisely measure and subsequently mitigate such algorithmic bias.

This project will examine the state of algorithmic (un)fairness within biomedical informatics. The SUDS Scholar will participate in a thorough literature review as well as several real data analyses (eg. EHRs, medical images) that illustrate the nuances of both identifying and measuring algorithmic bias in healthcare applications. The student will contribute to a manuscript for publication as well as open source software development.