Skip to main content Skip to secondary navigation
Main content start

AIMI Research Meeting: Attributing Performance Disparities in Medical AI to Distribution Shifts in Metadata - Kevin Wu, PhD Candidate

Event Details:

Thursday, October 20, 2022
3:00pm - 4:00pm PDT

Abstract

Medical AI algorithms often perform worse when deployed at different hospitals, hindering their widespread adoption. Such disparities are often attributed to shifts in underlying data distributions between sites, but subtle changes in high-dimensional inputs such as images or video are difficult to interpret and act upon. In this study, we instead examine performance disparities through the lens of distribution shifts in metadata, such as patient demographics, disease severity, and image acqusition methods. Such metadata are commonly collected in hospital systems and can be valuable to A) model developers in sourcing data for fine-tuning, and B) regulators and hospitals seeking to understand the sources of bias in an algorithm. Furthermore, as explanatory factors can often be confounded, (for example, disease severity and patient age), we propose a principled framework (ShapShift) for computing the marginal contribution of individual factors to overall performance disparities. We conduct a case-study of an algorithm trained to detect pneumothorax in chest X-rays and explain model degradation across multiple sites using a variety of metadata. Interestingly, we find that the majority of attributable drops in performance are due to changes in image acquisition as opposed to patient demographics, suggesting that technical factors should be a key consideration when addressing model biases.

About

Kevin Wu is a 3rd year PhD student at Stanford University in the Biomedical Informatics department. His research focuses on deployment of AI in medicine, including a focus on regulation and cross-site evaluation.

Explore More Events