New paper w/ @oziadias, @m_sendhil, and Christine Vogeli out today in @sciencemagazine.
Does a predictive algorithm widely used for population health management exhibit racial bias?
Short answer: YES.
Full Article: science.sciencemag.org/content/366/64…
THREAD (1/8)
cc: @sacjai @ashishkjha
Does a predictive algorithm widely used for population health management exhibit racial bias?
Short answer: YES.
Full Article: science.sciencemag.org/content/366/64…
THREAD (1/8)
cc: @sacjai @ashishkjha
BACKGROUND
Care management programs—which provide high-need, complex patients with additional services and support—have proliferated. An entire industry now exists to supply delivery organizations with predictive algorithms to help identify patients for these programs. (2/8)
Care management programs—which provide high-need, complex patients with additional services and support—have proliferated. An entire industry now exists to supply delivery organizations with predictive algorithms to help identify patients for these programs. (2/8)
WHAT WE DID
We looked for evidence of racial bias in a predictive algorithm used to identify patients for care management programs. This algorithm generates a “risk score” which prioritizes patients for increased services. It is one of the most widely used.(3/8)
We looked for evidence of racial bias in a predictive algorithm used to identify patients for care management programs. This algorithm generates a “risk score” which prioritizes patients for increased services. It is one of the most widely used.(3/8)
WHAT WE FOUND
At a given risk score, Black patients are considerably sicker than White patients (measured by medical complexity and signs of poorly controlled illness). For example, plot below shows diabetes control (y axis) and risk score percentile (x axis).(4/8)
At a given risk score, Black patients are considerably sicker than White patients (measured by medical complexity and signs of poorly controlled illness). For example, plot below shows diabetes control (y axis) and risk score percentile (x axis).(4/8)
WHY IT MATTERS
Since the risk score is used to prioritize patients for care management, this means fewer Black patients are prioritized. We conduct a simulation to show that eliminating this disparity would increase the % of Black patients getting extra help from 18 to 47.(5/8)
Since the risk score is used to prioritize patients for care management, this means fewer Black patients are prioritized. We conduct a simulation to show that eliminating this disparity would increase the % of Black patients getting extra help from 18 to 47.(5/8)
HOW DID THIS HAPPEN?
This algorithm (and most others) predict cost as a proxy for overall complexity/need (b/c its simpler and easier). But, due to unequal access and other structural inequalities, Black patients cost less than similarly (un)healthy White patients.(6/8)
This algorithm (and most others) predict cost as a proxy for overall complexity/need (b/c its simpler and easier). But, due to unequal access and other structural inequalities, Black patients cost less than similarly (un)healthy White patients.(6/8)
WHAT TO DO ABOUT IT
We're working with the algorithm developer to refocus the model on a more holistic picture of medical complexity+health needs, rather than just cost. We show in the paper that this approach still does a great job predicting cost—just without the bias.(7/8)
We're working with the algorithm developer to refocus the model on a more holistic picture of medical complexity+health needs, rather than just cost. We show in the paper that this approach still does a great job predicting cost—just without the bias.(7/8)
Thanks to @NIHCMfoundation for supporting this work! (8/8)
