Title: (Counterfactually) fair and accurate risk assessment for healthcare decision making
Authors: Alan Mishler - J.P. Morgan Chase (United States) [presenting]
Abstract: Algorithmic tools are increasingly used in healthcare settings to identify high-risk patients and inform treatment decisions, but these tools may perform differently across different racial groups (for example), leading to concerns about fairness. Although many methods exist for developing fair predictors, most such methods are concerned with observable outcomes, such as actual patient survival, which depends both on a patient's health and on the treatment they receive. By contrast, the accuracy and fairness properties of risk assessment tools are often most sensibly understood in terms of counterfactual outcomes: how a patient would fare if given, or not given, a particular treatment, irrespective of the treatment they actually receive. We (1) illustrate how a reliance on observable outcomes in risk assessment tools can worsen these outcomes, leading for example to higher patient mortality; and (2) describe a set of methods for building predictors that are both fair and accurate with respect to relevant counterfactual outcomes. These methods accommodate a wide range of fairness criteria, and they facilitate computationally efficient exploration of fairness-accuracy and fairness-fairness tradeoffs. In some cases, multiple unfairness measures can be simultaneously minimized with little cost in accuracy relative to a benchmark model.