Fairness in health AI: ensuring equity in healthcare cost.
summary
Walks through a fairness audit on a 2.4 million‑member claims set using Equitas to predict top 2% high‑cost patients, with demographics 90% White and 9.3% Black. Uses binary high‑cost label, demographic splits and recall/true positive rate as the chosen fairness metric. Finds predicted prevalence mirrors population share, predicted high‑cost ~2% while actual ~1.5%, and recall is higher for Black (0.43) than White (0.35), prompting metric choice like recall parity.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:48We are a health services innovation company
- 1:59We Connect and Serve Across the Health Care System
- 2:36General Principles in Ethical Al
- 5:00Healthcare Al Opportunities and Faimess
- 7:40Algorithmic Bias in The Context of Al and Health Systems
- 8:47Bias Points in ML Lifecycle
- 12:01Interrelated Factors Affect Health Disparities
- 16:42Cases of Healthcare Disparities in Minority Groups
- 18:56Disparities in Healthcare Data
- 21:41Proxies in Healthcare Data
- 22:59Choice of Model Evaluation Metric
- 25:29Tools for Evaluating ML Faimess
related talks
31mMeasuring Model Fairness
28mMeasuring Model Fairness
35mOpen Source Fairness
23mRemoving Unfair Bias in the Machine Learning Pipeline
18mU.S. Healthcare with AI: Cost Transparency and Affordability
11mAddressing Privacy, Bias, and Fairness in AI Systems
40mEthical AI: Addressing Bias & Algorithmic Fairness
50mThe Measure and Mismeasure of Fairness