On the interplay of accuracy and fairness in computational healthcare

Описание: Omer Reingold (Stanford University)
https://simons.berkeley.edu/talks/ome...
Theory of Computing and Healthcare

Machine-learning models are now routinely used to guide clinical decisions, allocate scarce resources, and assess patient risk. These systems raise well-motivated concerns about fairness, especially when performance varies across demographic or clinically meaningful subgroups. “Fairness” and “accuracy” are often framed as competing objectives in which social goals must come at the expense of predictive performance. Yet contemporary research in algorithmic fairness shows that this tradeoff does not hold for every definition of fairness.

In this talk, we will explore how fairness notions such as multicalibration and related indistinguishability-based definitions can, in fact, improve the reliability and robustness of predictive models. Multicalibration bridges the gap between actuarial (group-level) and clinical (individual-level) risk analysis by requiring predictions to be statistically valid across rich families of potentially overlapping subpopulations. We will discuss how these guarantees lead to models that are robust to distribution shifts and adaptable to changing downstream objectives and constraints without retraining.

Finally, we will highlight specific scenarios where fairness requirements do introduce genuine tensions with predictive accuracy and discuss why such tradeoffs may be unavoidable and raise policy decisions that healthcare systems must grapple with.