Domain: Healthcare/Medical AI
Description: Obermeyer et al. (2019) study in Science revealed racial bias in commercial health risk algorithms affecting \~200 million Americans annually, used to identify patients needing additional care programs.

Ethical Challenges:

• Measurement Bias: Algorithm used healthcare spending as proxy for health needs, but Black patients spent \~$1,800 less annually than white patients with identical conditions
• Disparate Impact: Black patients were 2.6 times less likely to receive additional care despite greater medical need
• Scale of Impact: Affects millions of healthcare decisions
• Proxy Problem: Seemingly neutral metrics can incorporate systemic bias

Public Datasets:

• Primary: MIMIC-CXR Database (PhysioNet)
• URL: https://physionet.org/content/mimic-cxr/2.0.0/
• Content: 377,110 chest X-ray images from Beth Israel Deaconess Medical Center
• Bias Documentation: Documented bias in medical image classifiers by race, sex, age, and insurance type
• Access: Requires credentials but publicly available for research