Ziad Obermeyer, M.D.
Blue Cross of California Distinguished Associate Professor of Health Policy and Management at the UC Berkeley School of Public Health
Asking Algorithms (and Ourselves) the Right Questions
Tuesday, May 4, 2021
7 p.m. | Online
Algorithms are moving out of the lab and into society, where they are asked to answer tough questions: which inmates to release, which job applicants to hire, which patients to treat. These questions cause algorithms to stumble, not so much because they are hard to answer, but because they are hard to ask. Humans easily grasp the idea of finding 'the sickest patients' or 'the best applicants'. Algorithms, by contrast, are incredibly literal: they can only deal with a particular variable in a particular dataset. Research is now showing that the specific ways in which abstract ideas get translated into machine-answerable questions can distort algorithms, leading them to encode bias and error.
In his lecture, Dr. Obermeyer will discuss one example of this from his own work, where a health algorithm created large-scale racial bias for tens of millions of patients, because it used health care costs as a proxy for health care needs. He will also give two reasons to be optimistic about the future of algorithms. First, many biases are fixable, by paying close attention to seemingly small technical choices when building algorithms, as we demonstrated by collaborating with the company that made the biased algorithm. Second, by forcing us to write down exactly what we mean -- what question we are asking exactly -- algorithms can hold up a mirror to our own biases, allowing us to understand and correct them.
Ziad Obermeyer, M.D. is the Blue Cross of California Distinguished Associate Professor of Health Policy and Management at the UC Berkeley School of Public Health, where he does research at the intersection of machine learning, medicine, and health policy. Dr. Obermeyer also continues to practice emergency medicine at the Fort Defiance Indian Hospital, Fort Defiance, AZ, Navajo Nation.