Fall2025ResearchProject

The deployment of medical AI systems across different hospitals raises critical questions about whether fairness and representation quality can be reliably transferred across clinical domains. Models trained on one hospital’s imaging data are often reused in new environments where patient demographics, imaging devices, and diagnostic practices differ substantially, potentially resulting in unintended bias against certain groups. This project investigates this challenge by studying fairness-aware representation alignment in medical imaging. The student will train contrastive learning models—such as SimCLR—independently on two large-scale chest X-ray datasets: CheXpert (from Stanford Hospital) and MIMIC-CXR (from Beth Israel Deaconess Medical Center). After learning embeddings in each domain, the student will apply domain alignment techniques such as Procrustes alignment to map representations from the CheXpert embedding space into the MIMIC-CXR space. The aligned embeddings will then be evaluated using fairness metrics designed for representation spaces, including demographic subgroup alignment, intra- vs. inter-group embedding disparity, and cluster-level demographic parity. The expected outcome is a rigorous understanding of whether fairness properties learned in one hospital setting preserve, degrade, or improve when transferred to another, revealing how robust model fairness is to realworld clinical domain shifts. A practical use case involves a healthcare network seeking to deploy a model trained at a major academic hospital (e.g., Stanford) into a community hospital setting: this project helps determine whether the transferred representations remain equitable across patient groups such as older adults, women, or specific disease cohorts. The findings will support responsible AI deployment in healthcare by highlighting the conditions under which fairness is stable across institutions and identifying scenarios where domain-specific mitigation strategies may be required.