Ai+ Training

Over the last two decades, the influence of machine learning on daily life has become increasingly common.  While the benefits of these applications are wide-reaching, notable high-profile cases of machine learning bias have raised concerns that algorithms will serve to widen societal inequality.   

The machine learning community has responded with a wide array of techniques aimed at measuring, and ultimately, eliminating such biases in machine learning models. Concerns about ML bias are particularly acute within the field of healthcare.  Healthcare, like many aspects of society, has underlying biases along lines of race, gender, and socioeconomic status.  These inequalities articulate themselves within the data used to build population health models.  If special care is not taken, it is all too easy for algorithms to guide population health efforts based on these inequalities under the guise of objectivity.     

A particular challenge is that canonical examples of bias mitigation have key distinguishing factors from typical health population application models, and techniques standard to ensuring model fairness are often inappropriate.  Certain bias mitigation approaches, taken by well-meaning practitioners, can harm communities that these techniques are designed to protect.  

The purpose of this talk is to introduce a novel measure of algorithmic fairness specifically designed for population health algorithms.  The discussion will begin by giving a brief introduction to ML bias and fairness techniques.  We’ll contrast the application of canonical examples used to illustrate ML bias with population health models, and contextualize how the techniques of bias mitigation fall short. 

We will then introduce our novel measure, called Group Benefit Equality, and explore a case study where it is used as a part of a larger effort to assess bias and fairness within models built for the CMS AI for Health Outcomes Challenge.    

The speaker has prepared an informal survey to gauge the audience's background in ML fairness. Please take 1 minute to answer this (very) short survey: https://www.surveymonkey.com/r/D9QB2T7