PowerTalks Seminar Series - Jin Chen, Ph.D. (Oct 13, 2022)

Topic: "Unbiased Representation of Clinical Data for Precise Patient Outcome Prediction"
Fairness is one of the newly emerging focuses for building trustworthy artificial intelligence (AI) models in healthcare. The behavior of a biased model often results in two problems: it performs significantly worse in certain populations than others, and it makes inequitable decisions towards patients with different backgrounds.
As a result, leaving the fairness problem unresolved might have a significant negative impact on deploying AI models in healthcare. Considering domain-specific and domain-invariant data representations as two special cases of data bias mitigation, we have developed a general, self-adaptive clinical data encoder model to learn unbiased and fair data representations from real-world Electronic Health Records (EHR).
The same fairness problem is also present in medical images, where image acquisition using non-standardized protocols often results in inconsistent image features, including texture, shape, or intensity. We have developed AI algorithms to harmonize medical images acquired using different acquisition protocols of the same scanner or images acquired using a similar protocol but with scanners from different vendors.